Now in Foundry: Tongyi-MAI Z-Image-Turbo, with FLUX.1-schnell and SDXL base 1.0
This week's Model Mondays edition pairs three models available through the Hugging Face collection in Microsoft Foundry: Tongyi-MAI's Z-Image-Turbo, a new designed for lower latency on a single GPU and native bilingual text rendering; Black Forest Labs' FLUX.1-schnell, a 12B rectified flow transformer distilled to 1–4 step inference and one of the most adopted open-weight image models since its 2024 release; and Stability AI's stable-diffusion-xl-base-1.0 (SDXL), a latent diffusion research model that can be used to generate and modify images based on text prompts. Models of the week Tongyi-MAI: Z-Image-Turbo Model Specs Parameters / size: 6B (BF16) Resolution: Up to 1024×1024 native Primary task: Text-to-image generation (English and Chinese) Why it's interesting (Spotlight) Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture: Z-Image concatenates text tokens, visual semantic tokens, and image VAE tokens into a single unified input stream rather than running text and image through separate branches. This single-stream design can improve parameter efficiency relative to dual-stream DiT architectures at the same capacity. See the Z-Image technical report for details. 8-step inference at sub-second latency, fits in 16GB VRAM: Z-Image-Turbo is distilled with Decoupled Distribution Matching Distillation (Decoupled-DMD) and further refined with DMDR, a method that fuses DMD with reinforcement learning during post-training. The result is a model that runs 8 Number-of-Function-Evaluations (NFE) per image with no Classifier-Free Guidance (CFG)—which roughly halves the per-step compute compared to CFG-based inference. See the Decoupled-DMD and DMDR papers. Native bilingual text rendering and strong instruction adherence: Unlike most open-weight image models, which struggle with legible in-image text, Z-Image-Turbo renders complex English and Chinese text accurately which is useful for posters, signage, packaging mockups, and marketing creative. Try it Imagine you're a community programs coordinator at your city's parks department, planning a new summer event series — a "Cake Picnic in the Park" — designed to bring neighbors together over food in shared green space. The event is a few weeks out. You haven't booked bakery partners yet, so no actual cake exists, and you need marketing assets this week to start driving sign-ups: a hero image for the registration page, a flyer for community centers and libraries, social tiles for the city's channels. Use the prompt below and a photorealistic image, that can now be scaled to become additional assets like printed flyers or social images in minutes using image editing tools (or another model). Prompt: A round layered cake displayed on a white ceramic cake stand, topped with glossy fresh red cherries and smooth pastel pink buttercream frosting piped in delicate rosettes around the edge. One generous slice has been cleanly cut and removed from the front, revealing a perfect cross-section: four distinct horizontal layers alternating between soft pink sponge cake and fluffy white vanilla cream frosting. Professional bakery photography, soft natural window light from the left, shallow depth of field, marble countertop, warm and inviting atmosphere, photorealistic detail on the cake texture, cherry highlights, and frosting swirls. Black Forest Labs: FLUX.1-schnell Model Specs Parameters / size: 12B (rectified flow transformer) Resolution: Flexible up to 2 megapixels Primary task: Text-to-image generation Why it's interesting (Spotlight) Rectified flow transformer with adversarial distillation for 1–4 step inference: FLUX.1-schnell is the distilled, Apache 2.0 sibling of the FLUX.1 family. It uses a rectified flow formulation (a diffusion variant that learns straight-line probability paths between noise and data, reducing the number of solver steps needed) and is further compressed with latent adversarial diffusion distillation. The model generates high quality images in for latency-sensitive workloads. Permissive licensing for commercial use: Released under Apache 2.0, FLUX.1-schnell can be used for personal, scientific, and commercial purposes. This has driven broad adoption across product features that need an open, redistributable image backbone. Strong prompt adherence at its parameter range: At 12B parameters, FLUX.1-schnell sits between the SDXL family and frontier proprietary image models, and it remains a common reference point for evaluating open image generation prompt following—particularly for complex compositional prompts and longer captions—roughly two years after its initial release. Try it Hugging Face Spaces give developers the ability to experiment and try new models before deploying them. Test out a few prompts here: https://black-forest-labs-flux-1-schnell.hf.space then when you are ready, deploy the model in Microsoft Foundry. Stability AI: stable-diffusion-xl-base-1.0 stabilityai/stable-diffusion-xl-base-1.0 · Hugging Face Model Specs Parameters / size: 2.6B UNet (≈3.5B total with text encoders) Resolution: 1024×1024 native Primary task: Text-to-image generation Why it's interesting (Spotlight) Dual text encoder design and an ensemble-of-experts pipeline: SDXL uses two pretrained text encoders—OpenCLIP-ViT/G and CLIP-ViT/L—concatenated to capture both broad semantic alignment and finer-grained token-level cues. It can be run standalone or paired with the SDXL refiner in an ensemble-of-experts pipeline where the base model handles early denoising and the refiner specializes in the final steps. See the SDXL report for the original training and architecture details. CreativeML Open RAIL++-M licensing for managed deployments: SDXL is distributed under the CreativeML Open RAIL++-M license, which permits commercial use and downstream fine-tuning with documented use restrictions. Try it To go deeper on SDXL, take a look at Stability AI's generative-models GitHub repository, which implements the most popular diffusion frameworks for both training and inference and continues to expand with new capabilities like distillation. Getting started You can deploy open-source Hugging Face models directly in Microsoft Foundry in two ways. The first by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. The second way is direct through the Hugging Face Hub, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure. Learn how to discover models and deploy them using Microsoft Foundry documentation: Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryNow in Foundry: IBM Granite 4.1, NVIDIA Nemotron Nano Omni, and Qwen3.6-35B-A3B
This week Microsoft Foundry adds two major model families alongside a reasoning powerhouse that spans the full spectrum from specialized speech and vision to general-purpose coding and long-context analysis. IBM's Granite 4.1 is a famiily of 10: six LLMs across 3B, 8B, and 30B sizes in both full-precision and FP8 variants, plus a safety model, a vision-language model for document extraction, and a multilingual speech recognition model. NVIDIA's Nemotron-3-Nano-Omni-30B-A3B-Reasoning brings multimodal capability—video, audio, image, and text—to a 31B Mamba2-Transformer Hybrid Mixture-of-Experts (MoE) architecture that activates only 3B parameters per forward pass; three variants are available in Foundry (BF16, FP8, and NVFP4), with the FP8 variant featured here. Qwen3.6-35B-A3B is designed for agentic coding among open models, with thinking preservation across conversation turns and a context window extensible to 1 million tokens. Models of the week IBM: granite-4.1-30b Model Specs Parameters / size: 30B (flagship of the Granite 4.1 family) Context length: 131,072 tokens Primary task: Text generation (multilingual instruction following, RAG, tool calling, code, summarization) Why it's interesting The Granite 4.1 release brings 10 models to Microsoft Foundry. The LLM lineup covers granite-4.1-3b-instruct, granite-4.1-8b-instruct, and granite-4.1-30b-instruct with FP8 variants for each, plus granite-guardian-4.1-8b for safety, granite-vision-4.1-4b for document and chart understanding, and granite-speech-4.1-2b for multilingual speech recognition. This is a deployment-ready stack where teams can mix and match model sizes and modalities from a single provider. Strong instruction following and reasoning at the 30B scale: granite-4.1-30b-instruct scores 80.16 on MMLU, 64.09 on MMLU-Pro, 83.74 on Big-Bench Hard (BBH), 77.80 on AGI Eval, 45.76 on GPQA (Graduate-Level Google-Proof Q&A, a graduate-level science reasoning benchmark), and 89.65 average on IFEval (instruction following). These results reflect SFT and reinforcement learning post-training focused specifically on instruction compliance, tool calling accuracy, and long-context retrieval. (View benchmarks here) Enhanced tool calling and 12-language support: Granite 4.1 models are trained for structured function calling and support 12 languages—Arabic, Chinese, Czech, Dutch, English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish—with dialog, extraction, and summarization capabilities. Safety and multimodal coverage within the same family: The inclusion of granite-guardian-4.1-8b (a safety classifier for detecting harmful content and prompt injections), granite-vision-4.1-4b (a Vision Language Model optimized for document extraction from PDFs, charts, and tables), and granite-speech-4.1-2b (a 2B multilingual Automatic Speech Recognition model) means teams can address safety, document parsing, and audio ingestion within the same model family—reducing integration complexity across a full pipeline. Try it Use Case Prompt Pattern Multilingual RAG Submit retrieved document passages in any of 12 supported languages; ask model to synthesize and cite sources Agentic tool calling Provide function definitions + user goal; model plans and executes tool calls in structured format Document extraction (granite-vision-4.1-4b) Submit PDF page image; extract tables, key figures, or form fields as structured JSON Safety classification (granite-guardian-4.1-8b) Pass user input or model output; receive structured risk assessment before serving response Sample prompt for an enterprise document processing deployment: You are building a multilingual document intelligence pipeline for a global financial institution. Using the granite-4.1-30b-instruct endpoint deployed in Microsoft Foundry, submit each incoming policy or regulatory document with the following system instruction: "You are a compliance analysis assistant. Review the document and extract: (1) all regulatory requirements described, (2) the entities to which each requirement applies, (3) any compliance deadlines mentioned, and (4) any penalties or consequences for non-compliance. Return the output as a structured JSON array with one entry per requirement." For documents that include scanned pages, first route them through granite-vision-4.1-4b to extract text and table content before passing to the 30B model for compliance analysis. Pass all user-facing outputs through granite-guardian-4.1-8b to screen for sensitive information before returning results. NVIDIA: Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 Model Specs Parameters / size: 31B total, ~3B activated per forward pass (Mamba2-Transformer Hybrid Mixture-of-Experts) Context length: 256,000 tokens Primary task: Video-audio-image-text-to-text (Multimodal understanding, reasoning, tool calling) Why it's interesting Multimodal input from a single efficient endpoint: Nemotron-3-Nano-Omni-30B-A3B-Reasoning supports video (up to 2 minutes), audio (up to 1 hour), images (RGB), and text—all from a single model deployed in Microsoft Foundry. Three variants are available in Foundry: full-precision BF16, FP8, and NVFP4. Paper: Nemotron Nano Omni technical report. Strong results across vision, document, video, and audio benchmarks: With reasoning mode enabled, the model scores 82.8 on MathVista-MINI (visual math reasoning), 67.04 on OCRBenchV2-EN (document OCR), 63.6 on Charxiv Reasoning (chart understanding), 72.2 on Video MME (video Q&A), 74.52 on Daily Omni (video+audio omnimodal understanding), and 89.39 on VoiceBench (speech instruction following). On OSWorld (GUI agent benchmark measuring autonomous computer use), it scores 47.4—a notable result for a model at the 3B active parameter scale. (Please see above model cards for further benchmark data) Mamba2-Transformer Hybrid MoE for efficient long-context inference: The model's layers alternate between Mamba2 state-space blocks (which process sequences with linear rather than quadratic cost) and standard Transformer attention blocks, combined with Mixture-of-Experts feedforward layers. Only ~3B parameters are activated per token despite the 31B total, making the 256K context window practically usable at lower compute cost than a comparably sized dense model. Word-level timestamps, JSON output, and tool calling for structured media workflows: The model produces word-level timestamps from audio, enabling precise transcript-to-timecode alignment for review and indexing workflows. Combined with JSON-structured output, chain-of-thought reasoning, and native tool calling, it can serve as an agentic step that ingests raw media (meeting recordings, M&E assets, training videos) and produces structured outputs for downstream systems without requiring separate transcription or OCR preprocessing stages. Try it Use Case Prompt Pattern Meeting intelligence Submit audio recording (up to 1 hr); extract transcript with word-level timestamps, action items, and decisions as structured JSON Video content analysis Submit video clip (up to 2 min) + query; retrieve timestamped summary of key events or spoken content Document + audio joint analysis Submit scanned document image alongside narrated walkthrough audio; extract and reconcile information from both modalities Multimodal tool calling Provide tool definitions + combined image/audio input; model reasons over content and executes structured tool calls Sample prompt for a media and compliance deployment: You are building a broadcast compliance review system for a media company. Using the Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 endpoint deployed in Microsoft Foundry, submit each recorded segment as video input with the following instruction: "Review this video segment and produce a compliance report as a JSON object with the following fields: transcript (full text with word-level timestamps), flagged_segments (array of objects with start_time, end_time, content, and reason for flagging), speaker_count (estimated number of distinct speakers), and compliance_summary (overall assessment). Flag any content that includes unverified factual claims, restricted product categories, or regulatory disclosures that may be incomplete." Use the word-level timestamps from the compliance report to route flagged segments directly to the editorial review queue with precise timecode references. Qwen: Qwen3.6-35B-A3B Model Specs Parameters / size: 35B total, 3B activated (Mixture-of-Experts) Context length: 262,144 tokens natively, extensible to 1,010,000 tokens Primary task: Image-text-to-text (agentic coding, reasoning, vision) Why it's interesting Agentic coding improvements over Qwen3.5-35B-A3B: Qwen3.6-35B-A3B scores 73.4 on SWE-bench Verified (vs. 70.0 for Qwen3.5-35B-A3B and 52.0 for Gemma 4 31B), 67.2 on SWE-bench Multilingual (vs. 60.3 and 51.7), and 49.5 on SWE-bench Pro (vs. 44.6 and 35.7). Terminal-Bench 2.0 reaches 51.5 (vs. 40.5 and 42.9). The update targets frontend workflows and repository-level reasoning specifically, areas where earlier Qwen3.5 iterations showed gaps. Blog post: Qwen3.6-35B-A3B. Hybrid architecture: Gated DeltaNet and Mixture-of-Experts: The model's 40 layers alternate between Gated DeltaNet blocks (a form of linear attention that avoids the quadratic cost of standard self-attention), Gated Attention blocks (using Grouped Query Attention with 16 query heads and 2 key-value heads), and Mixture-of-Experts (MoE) feedforward layers with 256 experts (8 routed + 1 shared active per token). Only 3B parameters are activated per forward pass, keeping inference cost comparable to a 3B dense model while retaining the capacity of a 35B model for knowledge and specialization. Thinking preservation across conversation turns: Qwen3.6 introduces an option to retain reasoning context from previous messages in multi-turn conversations. In prior models, chain-of-thought traces were stripped between turns, requiring the model to re-derive context it had already reasoned through. With thinking preservation enabled, iterative coding workflows—such as debugging across multiple exchanges—benefit from accumulated reasoning without repeating earlier analysis. Natively extensible to 1 million token context: The 262K native context is already among the largest in open models at this size, and the architecture supports extension to 1,010,000 tokens. On GPQA Diamond (science reasoning), Qwen3.6-35B-A3B scores 86.0—above both Gemma 4 31B (84.3) and Qwen3.5-27B (85.5)—while matching Gemma 4 31B on MMLU Pro (85.2) and LiveCodeBench v6 (80.4 vs. 80.0). Try it Use Case Prompt Pattern Repository-level code change Provide repository structure + task description; model plans file edits and outputs unified diff Multi-turn iterative debugging Enable thinking preservation; submit failing test + code across multiple turns; accumulate reasoning context Frontend code generation Provide design spec or screenshot + existing codebase context; generate component implementation Long-document reasoning Submit technical specification (up to 262K tokens); ask model to identify ambiguities or implementation gaps Sample prompt for a software engineering deployment: You are building an automated code review and implementation assistant for a platform engineering team. Using the Qwen3.6-35B-A3B endpoint deployed in Microsoft Foundry, enable thinking preservation for multi-turn sessions. In the first turn, submit the repository file tree and a GitHub issue describing a required API endpoint change. Prompt the model: "Review the repository structure and describe your implementation plan, including which files need to change and why." In the second turn, submit the relevant source files and prompt: "Based on your earlier plan, implement the changes and produce a unified diff." In the third turn, submit the test suite and prompt: "Write additional unit tests for the new endpoint, covering edge cases identified in your reasoning." The thinking preservation feature ensures the model carries forward its understanding of the codebase across all three turns without re-explaining context. Getting started You can deploy open-source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation. Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryNow in Foundry: Microsoft Harrier and NVIDIA EGM-8B
This week's Model Mondays edition highlights three models that share a common thread: each achieves results comparable to larger leading models, as a result of targeted training strategies rather than scale. Microsoft Research's harrier-oss-v1-0.6b from achieves state-of-the-art results on the Multilingual MTEB v2 embedding benchmark at 0.6B parameters through contrastive learning and knowledge distillation. NVIDIA's EGM-8B scores 91.4 average IoU on the RefCOCO visual grounding benchmark by training a small Vision Language Model (VLM) with reinforcement learning to match the output quality of much larger models. Together they represent a practical argument for efficiency-first model development: the gap between small and large models continues to narrow when training methodology is the focus rather than parameter count alone. Models of the week Microsoft Research: harrier-oss-v1-0.6b Model Specs Parameters / size: 0.6B Context length: 32,768 tokens Primary task: Text embeddings (retrieval, semantic similarity, classification, clustering, reranking) Why it's interesting State-of-the-art on Multilingual MTEB v2 from Microsoft Research: harrier-oss-v1-0.6b is a new embedding model released by Microsoft Research, achieving a 69.0 score on the Multilingual MTEB v2 (Massive Text Embedding Benchmark) leaderboard—placing it at the top of its size class at release. It is part of the harrier-oss family spanning harrier-oss-v1-270m (66.5 MTEB v2), harrier-oss-v1-0.6b (69.0), and harrier-oss-v1-27b (74.3), with the 0.6B variant further trained with knowledge distillation from the larger family members. Benchmarks: Multilingual MTEB v2 Leaderboard. Decoder-only architecture with task-instruction queries: Unlike most embedding models that use encoder-only transformers, harrier-oss-v1-0.6b uses a decoder-only architecture with last-token pooling and L2 normalization. Queries are prefixed with a one-sentence task instruction (e.g., "Instruct: Retrieve relevant passages that answer the query\nQuery: ...") while documents are encoded without instructions—allowing the same deployed model to be specialized for retrieval, classification, or similarity tasks through the prompt alone. Broad task coverage across six embedding scenarios: The model is trained and evaluated on retrieval, clustering, semantic similarity, classification, bitext mining, and reranking—making it suitable as a general embedding backbone for multi-task pipelines rather than a single-use retrieval model. One endpoint, consistent embeddings across the stack. 100+ language support: Trained on a large-scale mixture of multilingual data covering Arabic, Chinese, Japanese, Korean, and 100+ additional languages, with strong cross-lingual transfer for tasks that span language boundaries. Try it Use Case Prompt Pattern Multilingual semantic search Prepend task instruction to query; encode documents without instruction; rank by cosine similarity Cross-lingual document clustering Embed documents across languages; apply clustering to group semantically related content Text classification with embeddings Encode labeled examples + new text; classify by nearest-neighbor similarity in embedding space Bitext mining Encode parallel corpora in source and target languages; align segments by embedding similarity Sample prompt for a global enterprise knowledge base deployment: You are building a multilingual internal knowledge base for a global professional services firm. Using the harrier-oss-v1-0.6b endpoint deployed in Microsoft Foundry, encode all internal documents—policy guides, project case studies, and technical documentation—across English, French, German, and Japanese. At query time, prepend the task instruction to each employee query: "Instruct: Retrieve relevant internal documents that answer the employee's question\nQuery: {question}". Retrieve the top-5 most similar documents by cosine similarity and pass them to a language model with the instruction: "Using only the provided documents, answer the question and cite the source document title for each claim. If no document addresses the question, say so." NVIDIA: EGM-8B Model Specs Parameters / size: ~8.8B Context length: 262,144 tokens Primary task: Image-text-to-text (visual grounding) Why it's interesting Preforms well on visual grounding compared to larger models even at its small size: EGM-8B achieves 91.4 average Intersection over Union (IoU) on the RefCOCO benchmark—the standard measure of how accurately a model localizes a described region within an image. Compared to its base model Qwen3-VL-8B-Thinking (87.8 IoU), EGM-8B achieves a +3.6 IoU gain through targeted Reinforcement Learning (RL) fine-tuning. Benchmarks: EGM Project Page. 5.9x faster than larger models at inference: EGM-8B achieves 737ms average latency. The research demonstrates that test-time compute can be scaled horizontally across small models—generating many medium-quality responses and selecting the best—rather than relying on a single expensive forward pass through a large model. Two-stage training: EGM-8B is trained first with Supervised Fine-Tuning (SFT) on detailed chain-of-thought reasoning traces generated by a proprietary VLM, then refined with Group Relative Policy Optimization (GRPO) using a reward function combining IoU accuracy and task success. The intermediate SFT checkpoint is available as nvidia/EGM-8B-SFT for developers who want to experiment with the intermediate stage. Addresses a root cause of small model grounding errors: The EGM research identifies that 62.8% of small model errors on visual grounding stem from complex multi-relational descriptions—where a model must reason about spatial relationships, attributes, and context simultaneously. By focusing test-time compute on reasoning through these complex prompts, EGM-8B closes the gap without increasing the underlying model size. Try it Use Case Prompt Pattern Object localization Submit image + natural language description; receive bounding box coordinates Document region extraction Provide scanned document image + field description; extract specific regions Visual quality control Submit product image + defect description; localize defect region for downstream classification Retail shelf analysis Provide shelf image + product description; return location of specified SKU Sample prompt for a retail and logistics deployment: You are building a visual inspection system for a logistics warehouse. Using the EGM-8B endpoint deployed in Microsoft Foundry, submit each incoming package scan image along with a natural language grounding query describing the region of interest: "Please provide the bounding box coordinate of the region this sentence describes: {description}". For example: "the label on the upper-left side of the box", "the barcode on the bottom face", or "the damaged corner on the right side". Use the returned bounding box coordinates to route each package to the appropriate inspection station based on the identified region. Getting started You can deploy open-source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation: Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryNow in Foundry: Cohere Transcribe, Nanbeige 4.1-3B, and Octen Embedding
This week's Model Mondays edition spans three distinct layers of the AI application stack: Cohere's cohere-transcribe, a 2B Automatic Speech Recognition (ASR) model that ranks first on the Open ASR Leaderboard across 14 languages; Nanbeige's Nanbeige4.1-3B, a compact 3B reasoning model that outperforms models ten times its size on coding, math, and deep-search benchmarks; and Octen's Octen-Embedding-0.6B, a lightweight text embedding model that achieves strong retrieval scores across 100+ languages and industry-specific domains. Together, these three models illustrate how developers can build full AI pipelines—from audio ingestion to language reasoning to semantic retrieval—entirely with open-source models deployed through Microsoft Foundry. Each operates in a different modality and fills a distinct architectural role, making this week's selection especially well-suited for teams assembling production-grade systems across speech, text, and search. Models of the week Cohere's cohere-transcribe-03-2026 Model Specs Parameters / size: 2B Primary task: Automatic Speech Recognition (audio-to-text) Why it's interesting Top-ranked on the Open ASR Leaderboard: cohere-transcribe-03-2026 achieves a 5.42% average Word Error Rate (WER) across 8 English benchmark datasets as of March 26, 2026—placing it first among open models. It reaches 1.25% WER on LibriSpeech Clean and 8.15% on AMI (meeting transcription), demonstrating consistent accuracy across both clean speech and real-world, multi-speaker environments. Benchmarks: Open ASR Leaderboard. 14 languages with a dedicated encoder-decoder architecture: The model uses a large Conformer encoder for acoustic representation extraction paired with a lightweight Transformer decoder for token generation, trained from scratch on 14 languages covering European, East Asian (Chinese Mandarin, Japanese, Korean, Vietnamese), and Arabic. Unlike general-purpose models adapted for ASR, this dedicated architecture makes it efficient without sacrificing accuracy. Long-form audio with automatic chunking: Audio longer than 35 seconds is automatically split into overlapping chunks and reassembled into a coherent transcript—no manual preprocessing required. Batched inference, punctuation control, and per-language configuration are all supported through the standard API. Try it Click on the window above, upload an audio file, and watch how quickly the model transcribes it for you. Or click the link to experiment with the Cohere Transcribe Space and record audio directly from your device. Use Case Prompt Pattern Meeting transcription Submit recorded audio with language tag; retrieve timestamped transcript per speaker turn Call center quality review Batch-process customer call recordings, extract transcript, pass to classification model Medical documentation Transcribe clinical encounters; feed transcript into summarization or structured note pipeline Multilingual content indexing Process podcasts or video audio in any of 14 supported languages; store as searchable text Sample prompt for a legal services deployment: You are building a contract negotiation assistant. A client submits a recorded audio of a 45-minute supplier negotiation call. Using the cohere-transcribe-03-2026 endpoint deployed in Microsoft Foundry, transcribe the call with punctuation enabled for the English audio. Once the transcript is available, pass it to a downstream language model with the following instruction: "Identify all pricing commitments, delivery deadlines, and liability clauses mentioned in this negotiation transcript. For each, note the speaker's position (client or supplier) and flag any terms that appear ambiguous or require legal review." Nanbeige's Nanbeige4.1-3B Model Specs Parameters / size: 3B Context length: 131,072 tokens Primary task: Text generation (reasoning, coding, tool use, deep search) Why it's interesting Reasoning performance that exceeds its size class: Nanbeige4.1-3B scores 76.9 on LiveCodeBench-V6, these results suggest that targeted post-training using Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on a focused dataset can yield improvements that scale-based approaches cannot replicate at equivalent parameter counts. Read the technical report: https://huggingface.co/papers/2602.13367. Strong preference alignment at the 3B scale: On Arena-Hard-v2, Nanbeige4.1-3B scores 73.2, compared to 56.0 for Qwen3-32B and 60.2 for Qwen3-30B-A3B—both significantly larger models. This indicates that the model's outputs consistently match human preference for response quality and helpfulness, not just accuracy on structured tasks. Deep-search capability previously absent from small general models: On xBench-DeepSearch-2505, Nanbeige4.1-3B scores 75—matching search-specialized small agents. The model can sustain complex agentic tasks involving more than 500 sequential tool invocations, a capability gap that previously required either specialized search agents or significantly larger models. Native tool-use support: The model's chat template and generation pipeline natively support tool call formatting, making it straightforward to connect to external APIs and build multi-step agentic workflows without additional scaffolding. Try it Use Case Prompt Pattern Code review and fix Provide failing test + stack trace; ask model to diagnose root cause and write corrected implementation Competition-style math Submit problem as structured prompt; use temperature 0.6, top-p 0.95 for consistent reasoning steps Agentic task execution Provide tool definitions as JSON + goal; let model plan and execute tool calls sequentially Long-document Q&A Pass full document (up to 131K tokens) with targeted factual questions; extract structured answers Sample prompt for a software engineering deployment: You are automating pull request review for a backend engineering team. Using the Nanbeige4.1-3B endpoint deployed in Microsoft Foundry, provide the model with a unified diff of a proposed code change and the following system instruction: "You are a senior software engineer reviewing a pull request. For each modified function: (1) summarize what the change does, (2) identify any edge cases that are not handled, (3) flag any security or performance regressions relative to the original, and (4) suggest a specific improvement if one is warranted. Format your output as a structured list per function." Octen's Octen-Embedding-0.6B Model Specs Parameters / size: 0.6B Context length: 32,768 tokens Primary task: Text embeddings (semantic search, retrieval, similarity) Why it's interesting Retrieval performance above larger proprietary models at 0.6B: On the RTEB (Retrieval Text Embedding Benchmark) public leaderboard, Octen-Embedding-0.6B achieves a mean task score of 0.7241—above voyage-3.5 (0.7139), Cohere-embed-v4.0 (0.6534), and text-embedding-3-large (0.6110), despite being a fraction of their parameter count. The model is fine-tuned from Qwen3-Embedding-0.6B via Low-Rank Adaptation (LoRA), demonstrating that targeted fine-tuning on retrieval-specific data can close the gap with larger embedding models. Vertical domain coverage across legal, finance, healthcare, and code: Octen-Embedding-0.6B was trained with explicit coverage of domain-specific retrieval scenarios—legal document matching, financial report Q&A, clinical dialogue retrieval, and code search including SQL. This makes it suitable for regulated-industry applications where generic embedding models tend to underperform on specialized terminology. 32,768-token context for long-document retrieval: The extended context window supports encoding entire legal contracts, earnings reports, or clinical case notes as single embeddings—removing the need to chunk long documents and re-aggregate scores at query time, which can introduce ranking errors. 100+ language support with cross-lingual retrieval: The model handles multilingual and cross-lingual retrieval natively, with strong coverage across languages including English, Chinese, and other major languages via its Qwen3-based architecture—practical for global enterprise applications that span multiple languages. Use Case Prompt Pattern Semantic search Encode user query and document corpus; rank documents by cosine similarity to query embedding Legal precedent retrieval Embed case briefs and query with legal question; retrieve most semantically relevant precedents Cross-lingual document search Encode multilingual document set; submit query in any supported language for cross-lingual retrieval Financial Q&A pipeline Embed earnings reports or filings; retrieve relevant passages to ground downstream language model responses Sample prompt for a global enterprise knowledge base deployment: You are building a clinical decision support tool. Using the Octen-Embedding-0.6B endpoint deployed in Microsoft Foundry, embed a corpus of 10,000 clinical case notes at ingestion time and store the resulting 1024-dimensional vectors in a vector database. At query time, encode an incoming patient presentation summary and retrieve the 5 most semantically similar historical cases. Pass the retrieved cases and the current presentation to a language model with the following instruction: "Based on these five similar cases and their documented outcomes, summarize the most common treatment approaches and flag any cases where the outcome differed significantly from the initial prognosis." Getting started You can deploy open-source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation: Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryNow in Foundry: NVIDIA Nemotron-3-Super-120B-A12B, IBM Granite-4.0-1b-Speech, and Sarvam-105B
This week's Model Mondays edition highlights three models now available in Hugging Face collection on Microsoft Foundry: NVIDIA's Nemotron-3-Super-120B-A12B, a hybrid Latent Mixture-of-Experts (MOE) model with 12B active parameters and context handling up to 1 million tokens; IBM Granite's Granite-4.0-1b-Speech, a compact Automatic Speech Recognition (ASR) and Automatic Speech Translation (AST) model that achieves a 5.52% average Word Error Rate (WER) at 280× real-time speed with runtime keyword biasing for domain adaptation; and Sarvam's Sarvam-105B, a 105B Mixture-of-Experts (MoE) model with 10.3B active parameters optimized for complex reasoning and 22 Indian languages, with comparable agentic performance compared to other larger proprietary models on web search and task-planning benchmarks. Models of the week NVIDIA Nemotron-3-Super-120B-A12B Model Specs Parameters / size: 120B total with 12B active Context length: Up to 1M tokens Primary task: Text generation (reasoning, agentic workflows, long-context tasks, tool use, RAG) Why it's interesting (Spotlight) Hybrid Latent MoE architecture with selective attention: Nemotron-3-Super combines interleaved Mamba-2 state-space layers and sparse MoE layers with a select number of full attention layers—a design called Latent MoE. Tokens are routed into a smaller latent space for computation, which improves accuracy per parameter while keeping only 12B parameters active at inference time. Multi-Token Prediction (MTP) heads where the model simultaneously predicts multiple upcoming tokens during training enable native speculative decoding, reducing time-to-first-token on long outputs without a separate draft model. Configurable reasoning mode: The model supports toggling extended chain-of-thought reasoning on or off via the chat template flag enable_thinking. This lets developers suppress the reasoning trace for latency-sensitive tasks while keeping it available for high-stakes or multi-step agentic use cases without loading a separate model. Sustained 1M-token context reliability: On RULER, the standard long-context evaluation suite, Nemotron-3-Super achieves 91.75% at 1M tokens. This makes it practical for full-document retrieval-augmented generation (RAG), long-form code analysis, and extended agentic sessions without chunking or windowing strategies. Try it Use cases Best practices Ultra‑long document ingestion & consolidation (e.g., end‑to‑end review of massive specs, logs, or multi‑volume manuals without chunking) Use the native 1M‑token context to avoid windowing strategies; feed full corpora in one pass to reduce stitching errors. Prefer default decoding for general analysis (NVIDIA recommends temperature≈1.0, top_p≈0.95) before tuning; this aligns with the model’s training and MTP‑optimized generation path. Leverage MTP for throughput (multi‑token prediction improves output speed on long outputs), making single‑pass synthesis practical at scale. Latency‑sensitive chat & tool‑calling at scale (e.g., high‑volume enterprise assistants where response time matters) Toggle reasoning traces intentionally via the chat template (enable_thinking on/off): turn off for low‑latency interactions; on for harder prompts where accuracy benefits from explicit reasoning. Use model‑recommended sampling for tool calls (many guides tighten temperature for tool use) to improve determinism while keeping top_p near 0.95. Rely on the LatentMoE + MTP design to sustain high tokens/sec under load instead of adding a draft model for speculative decoding. IBM Granite-4.0-1b-Speech Model Specs Parameters / size: ~1B Context length: 128K tokens (LLM backbone; audio processed per utterance through the speech encoder) Primary task: Multilingual Automatic Speech Recognition (ASR) and bidirectional Automatic Speech Translation (AST) Why it's interesting (Spotlight) Compact ASR with speculative decoding at near-real-time speed: At roughly 1B parameters, Granite-4.0-1b-Speech achieves a 5.52% average WER across eight English benchmarks at 280× real-time speed (RTFx—the ratio of audio duration processed to wall-clock time) on the Open ASR Leaderboard. Runtime keyword biasing for domain adaptation without fine-tuning: Granite-4.0-1b-Speech accepts a runtime keyword list—proper nouns, brand names, technical terms, acronyms—that adjusts decoding probabilities toward those terms. This allows domain-specific vocabulary to be injected at inference time rather than requiring a fine-tuning run, practical for legal transcription, medical dictation, or financial meeting notes where terminology changes across clients. Bidirectional speech translation across 6 languages in one model: Beyond ASR, the model supports translation both to and from English for French, German, Spanish, Portuguese, and Japanese, plus English-to-Italian and English-to-Mandarin. A single deployed endpoint handles ASR and AST tasks without routing audio to separate models, reducing infrastructure surface area. Try it Test the model in the Hugging Face space before deploying in Foundry here: Sarvam’s Sarvam-105B Model Specs Parameters / size: 105B total with 10.3B active (Mixture of Experts, BF16) Context length: 128K tokens (with YaRN-based long-context extrapolation, scale factor 40) Primary task: Text generation (reasoning, coding, agentic tasks, Indian language understanding) Why it's interesting (Spotlight) Broad Indian language coverage at scale: Sarvam-105B supports English and 22 Indian languages—Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, Urdu, Sanskrit, Maithili, Dogri, Manipuri, Santali, Kashmiri, Nepali, Sindhi, Konkani, and Tibetan—the broadest open-model coverage for this language set at this parameter range. Training explicitly prioritized the Indian context, resulting in reported state-of-the-art performance across these languages for models of comparable size. Strong agentic and web-search performance: Sarvam-105B scores 49.5% on BrowseComp (web research benchmark with search tool access)—substantially above GLM-4.5-Air (21.3%) and Qwen3-Next-80B-A3B-Thinking (38.0%). It also achieves 68.3% average on τ² Bench (multi-domain task-planning benchmark), above GPT-OSS-120B (65.8%) and GLM-4.5-Air (53.2%). This reflects training emphasis on multi-step agentic workflows in addition to standard reasoning. Try it Use cases Best practices Agentic web research & technical troubleshooting (multi-step reasoning, planning, troubleshooting) Use longer context when needed: the model is designed for long-context workflows (up to 128K context with YaRN-based extrapolation noted). Start from the model’s baseline decoding settings (as shown in the model’s sample usage) and adjust for your task: temperature ~0.8, top_p ~0.95, repetition_penalty ~1.0, and set an explicit max_new_tokens (sample shows 2048). Suggestion (general, not stated verbatim in the sources): For agentic tasks, keep the prompt structured (goal → constraints → tools available → required output format), and ask for a short plan + final answer to reduce wandering. Multilingual (Indic) customer support & content generation (English + 22 Indian languages; native-script / romanized / code-mixed inputs) Be explicit about the language/script you want back (e.g., Hindi in Devanagari vs romanized Hinglish), since training emphasized Indian languages and code-mixed/romanized inputs. Provide in-language examples (a short “good response” example in the target language/script) to anchor tone and terminology. (Suggestion—general best practice; not stated verbatim in sources.) Use the model’s baseline generation settings first (sample decoding params) and then tighten creativity for support use cases (e.g., lower temperature) if you see variability. Getting started You can deploy open-source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. Or start from the Hugging Face Hub and choose the "Deploy on Microsoft Foundry" option, which brings you straight into Foundry. Learn how to discover models and deploy them using Microsoft Foundry here: Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryNow in Foundry: VibeVoice-ASR, MiniMax M2.5, Qwen3.5-9B
This week's Model Mondays edition features two models that have just arrived in Microsoft Foundry: Microsoft's VibeVoice-ASR, a unified speech-to-text model that handles 60-minute audio files in a single pass with built-in speaker diarisation and timestamps, and MiniMaxAI's MiniMax-M2.5, a frontier agentic model that leads on coding and tool-use benchmarks with performance comparable to the strongest proprietary models at a fraction of their cost; and Qwen's Qwen3.5-9B, the largest of the Qwen3.5 Small Series. All three represent a shift toward long-context, multi-step capability: VibeVoice-ASR processes up to an hour of continuous audio without chunking; MiniMax-M2.5 handles complex, multi-phase agentic tasks more efficiently than its predecessor—completing SWE-Bench Verified 37% faster than M2.1 with 20% fewer tool-use rounds; and Qwen3.5-9B brings multimodal reasoning on consumer hardware that outperforms much larger models. Models of the week VibeVoice-ASR Model Specs Parameters / size: ~8.3B Primary task: Automatic Speech Recognition with diarisation and timestamps Why it's interesting 60-minute single-pass with full speaker attribution: VibeVoice-ASR processes up to 60 minutes of continuous audio without chunk-based segmentation—yielding structured JSON output with start/end timestamps, speaker IDs, and transcribed content for each segment. This eliminates the speaker-tracking drift and semantic discontinuities that chunk-based pipelines introduce at segment boundaries. Joint ASR, diarisation, and timestamps in one model: Rather than running separate systems for transcription, speaker separation, and timing, VibeVoice-ASR produces all three outputs in a single forward pass. Users can also inject customized hot words—proper nouns, technical terms, or domain-specific phrases—to improve recognition accuracy on specialized content without fine-tuning. Multilingual with native code-switching: Supports 50+ languages with no explicit language configuration required and handles code-switching within and across utterances natively. This makes it suitable for multilingual meetings and international call center recordings without pre-routing audio by language. Benchmarks: On the Open ASR Leaderboard, VibeVoice-ASR achieves an average WER of 7.77% across 8 English datasets (RTFx 51.80), including 2.20% on LibriSpeech Clean and 2.57% on TED-LIUM. On the MLC-Challenge multi-speaker benchmark: DER 4.28%, cpWER 11.48%, tcpWER 13.02%. Try it Use case What to build Best practices Long-form, multi-speaker transcription for meetings + compliance A transcription service that ingests up to 60 minutes of audio per request and returns structured segments with speaker IDs + start/end timestamps + transcript text (ready for search, summaries, or compliance review). Keep audio un-chunked (single-pass) to preserve speaker coherence and avoid stitching drift; rely on the model’s joint ASR, diarisation, and timestamping so you don’t need separate diarisation/timestamp pipelines or postprocessing. Multilingual + domain-specific transcription (global support, technical reviews) A global transcription workflow for multilingual meetings or call center recordings that outputs “who/when/what,” and supports vocabulary injection for product names, acronyms, and technical terms. Provide customized hot words (names / technical terms) in the request to improve recognition on specialized content; don’t require explicit language configuration—VibeVoice-ASR supports 50+ languages and code-switching, so you can avoid pre-routing audio by language. Read more about the model and try out the playground Microsoft for Hugging Face Spaces to try the model for yourself. MiniMax-M2.5 Model Specs Parameters / size: ~229B (FP8, Mixture of Experts) Primary task: Text generation (agentic coding, tool use, search) Why it's interesting? Leading coding benchmark performance: Scores 80.2% on SWE-Bench Verified and 51.3% on Multi-SWE-Bench across 10+ programming languages (Go, C, C++, TypeScript, Rust, Python, Java, and others). In evaluations across different agent harnesses, M2.5 scores 79.7% on Droid and 76.1% on OpenCode—both ahead of Claude Opus 4.6 (78.9% and 75.9% respectively). The model was trained across 200,000+ real-world coding environments covering the full development lifecycle: system design, environment setup, feature iteration, code review, and testing. Expert-level search and tool use: M2.5 achieves industry-leading performance in BrowseComp, Wide Search, and Real-world Intelligent Search Evaluation (RISE), laying a solid foundation for autonomously handling complex tasks. Professional office work: Achieves a 59.0% average win rate against other mainstream models in financial modeling, Word, and PowerPoint tasks, evaluated via the GDPval-MM framework with pairwise comparison by senior domain professionals (finance, law, social sciences). M2.5 was co-developed with these professionals to incorporate domain-specific tacit knowledge—rather than general instruction-following—into the model's training. Try it Use case What to build Best practices Agentic software engineering Multi‑file code refactors, CI‑gated patch generation, long‑running coding agents working across large repositories Start prompts with a clear architecture or refactor goal. Let the model plan before editing files, keep tool calls sequential, and break large changes into staged tasks to maintain state and coherence across long workflows. Autonomous productivity agents Research assistants, web‑enabled task agents, document and spreadsheet generation workflows Be explicit about intent and expected output format. Decompose complex objectives into smaller steps (search → synthesize → generate), and leverage the model’s long‑context handling for multi‑step reasoning and document creation. With these use cases and best practices in mind, the next step is translating them into a clear, bounded prompt that gives the model a specific goal and the right tools to act. The example below shows how a product or engineering team might frame an automated code review and implementation task, so the model can reason through the work step by step and return results that map directly back to the original requirement: “You're building an automated code review and feature implementation system for a backend engineering team. Deploy MiniMax-M2.5 in Microsoft Foundry with access to your repository's file system tools and test runner. Given a GitHub issue describing a new API endpoint requirement, have the model first write a functional specification decomposing the requirement into sub-tasks, then implement the endpoint across the relevant service files, write unit tests with at least 85% coverage, and return a pull request summary explaining each code change and its relationship to the original requirement. Flag any implementation decisions that deviate from the patterns found in the existing codebase.” Qwen3.5-9B Model Specs Parameters / size: 9B Context length: 262,144 tokens natively; extensible to 1,010,000 tokens Primary task: Image-text-to-text (multimodal reasoning) Why it’s interesting High intelligence density at small sizes: Qwen 3.5 Small models show large reasoning gains relative to parameter count, with the 4B and 9B variants outperforming other sub‑10B models on public reasoning benchmarks. Long‑context by default: Support for up to 262K tokens enables long‑document analysis, codebase review, and multi‑turn workflows without chunking. Native multimodal architecture: Vision is built into the model architecture rather than added via adapters, allowing small models (0.8B, 2B) to handle image‑text tasks efficiently. Open and deployable: Apache‑2.0 licensed models designed for local, edge, or cloud deployment scenarios. Benchmarks AI Model & API Providers Analysis | Artificial Analysis Try it Use case When to use Best‑practice prompt pattern Long‑context reasoning Analyzing full PDFs, long research papers, or large code repositories where chunking would lose context Set a clear goal and scope. Ask the model to summarize key arguments, surface contradictions, or trace decisions across the entire document before producing an output. Lightweight multimodal document understanding OCR‑driven workflows using screenshots, scanned forms, or mixed image‑text inputs Ground the task in the artifact. Instruct the model to first describe what it sees, then extract structured information, then answer follow‑up questions. With these best practices in mind, Qwen 3.5-9B demonstrates how compact, multimodal models can handle complex reasoning tasks without chunking or manual orchestration. The prompt below shows how an operations analyst might use the model to analyze a full report end‑to‑end: "You are assisting an operations analyst. Review the attached PDF report and extracted tables. Identify the three largest cost drivers, explain how they changed quarter‑over‑quarter, and flag any anomalies that would require follow‑up. If information is missing, state what data would be needed." Getting started You can deploy open-source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation. Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryNow in Foundry: Qwen3.5 Medium Model Series
This week's spotlight focuses on the Qwen3.5 Medium Model Series, now available in Microsoft Foundry. All three models are Vision Language Models (VLMs) built with early-fusion multimodal training, a 262K native context window, and support for 201 languages, released under Apache 2.0. They range from a 27B dense model optimized for latency-sensitive deployments to a 122B sparse Mixture-of-Experts (MoE) model that activates only 10B parameters per inference call, delivering frontier-class multimodal performance at lower inference cost. Models of the week What the Qwen3.5 Medium Model Series brings Before looking at each model individually, three architectural advances apply to all three and are worth understanding: Unified Vision-Language training (early fusion): Rather than attaching a separate vision encoder to a text model as an afterthought, Qwen3.5 trains on text and image tokens together from the beginning. This can enable stronger reasoning over diagrams, charts, and documents compared to prior Qwen3-VL models, which used a separate vision pipeline. Gated Delta Networks: A novel linear attention mechanism that replaces standard self-attention in most transformer layers. Combined with sparse MoE routing in the two larger models, this hybrid can deliver high-throughput inference at lower latency than equivalent dense architectures. Scalable RL across agent environments: Post-training uses reinforcement learning scaled across large multi-agent environments, contributing to strong performance on instruction-following and agentic task benchmarks. On vision-language reasoning tasks like MMMU and MathVista, these are models small enough to run on local hardware, yet competitive with large, frontier models on multimodal benchmarks. Qwen3.5-27B Model Specs Parameters / size: 27B (dense) Context length: 262,144 tokens Primary task: Vision Language Model (image-text-to-text) Why it's interesting (Spotlight) The dense baseline of the family: Unlike its MoE siblings, Qwen3.5-27B activates all 27B parameters on every forward pass. This gives it predictable, consistent latency per token—an important property for real-time applications and latency-sensitive deployments where MoE routing variability is a concern. Instruction-following leader across the family: Scores 95.0 on IFEval, the highest in the family (vs 93.4 for 122B-A10B and 91.9 for 35B-A3B), and 76.5 on IFBench—making it the strongest choice for structured-output tasks, complex multi-step instruction chains, and agent scaffolds that rely on precise format compliance. Try it You're building a visual quality inspection system for a circuit board manufacturer. Deploy Qwen3.5-27B in Microsoft Foundry to process images captured by a production line camera. Manufacturing sample prompt: Given an image of a printed circuit board (PCB), identify visible defects such as solder bridges, missing components, or misaligned pads. Return a JSON object with defect type, approximate board location, and severity (low / medium / high). Flag any board containing at least one high-severity defect for immediate rework routing. Qwen3.5-35B-A3B Model Specs Parameters / size: 35B total, 3B activated per forward pass (MoE) Context length: 262,144 tokens Primary task: Vision Language Model (image-text-to-text) Why it's interesting (Spotlight) The throughput-optimized pick: With only 3B parameters active per token despite a 35B parameter pool, this model delivers performance close to much larger dense models at substantially lower inference cost. 256-expert MoE routing at compact scale: Routes each token through 8 of 256 routed experts plus 1 shared expert. This breadth of specialization at a scale that only activates 3B parameters makes the 35B-A3B well-suited for high-throughput serving scenarios where cost per inference matters. Try it You're building a contract review assistant for an in-house legal team at a multinational company. Deploy Qwen3.5-35B-A3B in Microsoft Foundry to process scanned contract pages provided as images. Legal document sample prompt: Given a page from a commercial services agreement, extract all defined terms, identify obligation and liability clauses, and flag any termination conditions that deviate from standard commercial practice. Return a structured summary with clause type, section reference, and a one-sentence plain-language explanation of each flagged item. Qwen3.5-122B-A10B Model Specs Parameters / size: 122B total, 10B activated per forward pass (MoE) Context length: 262,144 tokens Primary task: Vision Language Model (image-text-to-text) Why it's interesting (Spotlight) Highest capability in the family: Leads across most benchmarks—76.9 on MMMU-Pro, 83.9 on MMMU, and 86.7 on MMLU-Pro. It also leads the family on SuperGPQA at 67.1 and MMLU-Redux at 94.0, reflecting stronger expert-level knowledge depth. Vision + language reasoning at scale: With the largest routing pool (256 experts, 8 routed + 1 shared) and 10B active parameters, this model handles the most demanding multimodal tasks in the family—long-document analysis over images, multi-step visual reasoning, and complex cross-modal instruction following at extended context lengths. Try it You're building an earnings research assistant for an investment team. Deploy Qwen3.5-122B-A10B in Microsoft Foundry to analyze earnings presentation slides submitted as images. Financial research sample prompt: Given a slide containing a combination of charts, tables, and management commentary, extract key financial metrics (revenue, EBITDA, year-over-year growth), interpret the trend shown in any charts, and generate a two-paragraph analyst summary suitable for a morning briefing. Flag any metrics that deviate materially from prior-quarter guidance and indicate the direction of the deviation. Getting started You can deploy open-source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation. Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryWhat’s trending on Hugging Face: PubMedBERT Base Embeddings, Paraphrase Multilingual MiniLM, BGE-M3
The embedding model landscape has evolved beyond one-size-fits-all solutions. Today’s developers navigate a set of deliberate trade‑offs: domain specialization to improve accuracy in vertical applications, multilingual capabilities to support global use cases, and retrieval strategies that optimize performance at scale. Once a model demonstrates strong semantic performance, predictable behavior, and broad community support, it often becomes a trusted reference baseline that developers build around and deploy with confidence. This week, we’re not spotlighting models that are new to Microsoft Foundry. Instead, we’re turning our attention to models that have managed to stay relevant in a rapidly expanding sea of options. This week's Model Monday's edition highlights three Hugging Face models including NeuML's PubMedBERT Base Embeddings for domain-specific medical text understanding, Sentence Transformers' Paraphrase Multilingual MiniLM for lightweight cross-lingual semantic similarity, and BAAI's BGE-M3 for multi-functional long-context retrieval across 100+ languages. Models of the week NeuML: PubMedBERT Base Embeddings Model Specs Parameters / size: 109M Context length: 512 tokens Primary task: Embeddings (medical domain) Why it's interesting Domain-specific performance gains: Fine-tuned on PubMed title-abstract pairs, achieving 95.62% average Pearson correlation across medical benchmarks—outperforming general-purpose models like gte-base (95.37%), bge-base-en-v1.5 (93.78%), and all-MiniLM-L6-v2 (93.46%) on medical literature tasks Production-validated for medical RAG: With 141K downloads and deployment in 30+ medical AI applications, this model demonstrates consistent real-world performance for clinical research, drug discovery, and biomedical semantic search pipelines Built on Microsoft's BiomedNLP foundation: Extends BioMed BERT family with sentence-transformers mean pooling, creating 768-dimensional embeddings optimized for medical literature clustering and retrieval Try it Clinical research sample prompt: Industry specific sample prompt: You're building a clinical decision support system for oncology. Deploy PubMedBERT Base Embeddings in Microsoft Foundry to index 50,000 recent cancer research abstracts from PubMed. A physician queries: "What are the cardiotoxicity risks of combining checkpoint inhibitors with anthracycline chemotherapy in elderly patients?" Embed the query, retrieve the top 10 most semantically similar abstracts using cosine similarity, and return citations with PubMed IDs for evidence-based treatment planning. Sentence Transformers: Paraphrase Multilingual MiniLM L12 v2 Model Specs Parameters / size: 117M Context length: 128 tokens Primary task: Embeddings (multilingual, sentence similarity) Why it's interesting Multilingual adoption: Supports 50+ languages including Arabic, Chinese, Hebrew, Hindi, Japanese, Korean, Russian, Thai, and Vietnamese—with 18.4 million downloads last month demonstrating production-scale validation across global deployments Compact architecture for edge deployment: At 117M parameters producing 384-dimensional embeddings, this model balances multilingual coverage with inference efficiency, making it ideal for resource-constrained environments or high-throughput applications Sentence-BERT foundation: Based on the influential Sentence-BERT paper (Reimers & Gurevych, 2019), using siamese BERT networks with mean pooling to create semantically meaningful sentence embeddings for clustering, paraphrase detection, and cross-lingual search Community-proven versatility: With 299 fine-tuned variants and 100+ Spaces implementations, this model serves as a peer reviewed starting point for multilingual semantic similarity tasks, from customer support ticket routing to cross-lingual document retrieval Try it E-commerce sample prompt: You're building a global customer support platform for an e-commerce company operating in 30 countries. Deploy Paraphrase Multilingual MiniLM in Microsoft Foundry to process incoming support tickets in English, Spanish, French, German, Portuguese, Japanese, and Korean. Embed each ticket as a 384-dimensional vector and cluster by semantic similarity to automatically route issues to specialized teams (payment, shipping, returns, technical). Flag duplicate tickets with cosine similarity > 0.85 to prevent redundant responses. BAAI: BGE-M3 Model Specs Parameters / size: ~560M Context length: 8192 tokens Primary task: Embeddings (multi-functional: dense, sparse, multi-vector) Why it's interesting Three retrieval modes in one model: Uniquely supports dense retrieval (1024-dim embeddings), sparse retrieval (lexical matching like BM25), and multi-vector retrieval (ColBERT-style fine-grained matching)—enabling hybrid search pipelines without maintaining separate models or indexes Exceptional long-context capability: 8192-token context window handles full documents, legal contracts, research papers, and lengthy technical content—validated on MLDR (13-language document retrieval) and NarrativeQA (long-form question answering) benchmarks Multilingual dominance: Outperforms OpenAI embeddings on MIRACL multilingual retrieval across 13+ languages and demonstrates strong zero-shot cross-lingual transfer on MKQA. Try it Legal document search sample prompt: You're building a legal document search system for a multinational law firm. Deploy BGE-M3 in Microsoft Foundry to index 5,000 full-length commercial contracts (average 6,000 tokens each) in English, French, German, and Spanish. A lawyer queries: "Find all force majeure clauses that exclude liability for pandemics or global health emergencies." Use hybrid retrieval: (1) dense embeddings for semantic similarity to capture concept variations like "Act of God" or "unforeseen circumstances", (2) sparse retrieval for exact keyword matches on "force majeure", "pandemic", "health emergency". Combine scores with weighted sum (0.6 dense + 0.4 sparse) and return top 15 contract sections with clause numbers and jurisdiction metadata. Getting started You can deploy open-source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation. Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryNow in Foundry: Qwen3-Coder-Next, Qwen3-ASR-1.7B, Z-Image
This week's spotlight features three models from that demonstrate enterprise-grade AI across the full scope of modalities. From low latency coding agents to state-of-the-art multilingual speech recognition and foundation-quality image generation, these models showcase the breadth of innovation happening in open-source AI. Each model balances performance with practical deployment considerations, making them viable for production systems while pushing the boundaries of what's possible in their respective domains. This week's Model Mondays edition highlights Qwen3-Coder-Next, an 80B MoE model that activates only 3B parameters while delivering coding agent capabilities with 256k context; Qwen3-ASR-1.7B, which achieves state-of-the-art accuracy across 52 languages and dialects; and Z-Image from Tongyi-MAI, an undistilled text-to-image foundation model with full Classifier-Free Guidance support for professional creative workflows. Models of the week Qwen: Qwen3-Coder-Next Model Specs Parameters / size: 80B total (3B activated) Context length: 262,144 tokens Primary task: Text generation (coding agents, tool use) Why it's interesting Extreme efficiency: Activates only 3B of 80B parameters while delivering performance comparable to models with 10-20x more active parameters, making advanced coding agents viable for local deployment on consumer hardware Built for agentic workflows: Excels at long-horizon reasoning, complex tool usage, and recovering from execution failures, a critical capability for autonomous development that go beyond simple code completion Benchmarks: Competitive performance with significantly larger models on SWE-bench and coding benchmarks (Technical Report) Try it Use Case Prompt Pattern Code generation with tool use Provide task context, available tools, and execution environment details Long-context refactoring Include full codebase context within 256k window with specific refactoring goals Autonomous debugging Present error logs, stack traces, and relevant code with failure recovery instructions Multi-file code synthesis Describe architecture requirements and file structure expectations Financial services sample prompt: You are a coding agent for a fintech platform. Implement a transaction reconciliation service that processes batches of transactions, detects discrepancies between internal records and bank statements, and generates audit reports. Use the provided database connection tool, logging utility, and alert system. Handle edge cases including partial matches, timing differences, and duplicate transactions. Include unit tests with 90%+ coverage. Qwen: Qwen3-ASR-1.7B Model Specs Parameters / size: 1.7B Context length: 256 tokens (default), configurable up to 4096 Primary task: Automatic speech recognition (multilingual) Why it's interesting All-in-one multilingual capability: Single 1.7B model handles language identification plus speech recognition for 30 languages, 22 Chinese dialects, and English accents from multiple regions—eliminating the need to manage separate models per language Specialized audio versatility: Transcribes not just clean speech but singing voice, songs with background music, and extended audio files, expanding use cases beyond traditional ASR to entertainment and media workflows State-of-the-art accuracy: Outperforms GPT-4o, Gemini-2.5, and Whisper-large-v3 across multiple benchmarks. English: Tedlium 4.50 WER vs 7.69/6.15/6.84; Chinese: WenetSpeech 4.97/5.88 WER vs 15.30/14.43/9.86 (Technical Paper) Language ID included: 97.9% average accuracy across benchmark datasets for automatic language identification, eliminating the need for separate language detection pipelines Try it Use Case Prompt Pattern Multilingual transcription Send audio files via API with automatic language detection Call center analytics Process customer service recordings to extract transcripts and identify languages Content moderation Transcribe user-generated audio content across multiple languages Meeting transcription Convert multilingual meeting recordings to text for documentation Customer support sample prompt: Deploy Qwen3-ASR-1.7B to a Microsoft Foundry endpoint and transcribe multilingual customer service calls. Send audio files via API to automatically detect the language (from 52 supported options including 30 languages and 22 Chinese dialects) and generate accurate transcripts. Process calls from customers speaking English, Spanish, Mandarin, Cantonese, Arabic, French, and other languages without managing separate models per language. Use transcripts for quality assurance, compliance monitoring, and customer sentiment analysis. Tongyi-MAI: Z-Image Model Specs Parameters / size: 6B Context length: N/A (text-to-image) Primary task: Text-to-image generation Why it's interesting Undistilled foundation model: Full-capacity base without distillation preserves complete training signal with Classifier-Free Guidance support (a technique that improves prompt adherence and output quality), enabling complex prompt engineering and negative prompting that distilled models cannot achieve High output diversity: Generates distinct character identities in multi-person scenes with varied compositions, facial features, and lighting, critical for creative applications requiring visual variety rather than consistency Aesthetic versatility: Handles diverse visual styles from hyper-realistic photography to anime and stylized illustrations within a single model, supporting resolutions from 512×512 to 2048×2048 at any aspect ratio with 28-50 inference steps (Technical Paper) Try it Use Case Prompt Pattern Multilingual transcription Send audio files via API with automatic language detection Call center analytics Process customer service recordings to extract transcripts and identify languages Content moderation Transcribe user-generated audio content across multiple languages Meeting transcription Convert multilingual meeting recordings to text for documentation E-commerce sample prompt: Professional product photography of a modern ergonomic office chair in a bright Scandinavian-style home office. Natural window lighting from left, clean white desk with laptop and succulent plant, light oak hardwood floor. Chair positioned at 45-degree angle showing design details. Photorealistic, commercial photography, sharp focus, 85mm lens, f/2.8, soft shadows. Getting started You can deploy open‑source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation. Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft FoundryWhat is trending in Hugging Face on Microsoft Foundry? Feb, 2, 2026
Open‑source AI is moving fast, with important breakthroughs in reasoning, agentic systems, multimodality, and efficiency emerging every day. Hugging Face has been a leading platform where researchers, startups, and developers share and discover new models. Microsoft Foundry brings these trending Hugging Face models into a production‑ready experience, where developers can explore, evaluate, and deploy them within their Azure environment. Our weekly Model Monday’s series highlights Hugging Face models available in Foundry, focusing on what matters most to developers: why a model is interesting, where it fits, and how to put it to work quickly. This week’s Model Mondays edition highlights three Hugging Face models, including a powerful Mixture-of-Experts model from Z. AI designed for lightweight deployment, Meta’s unified foundation model for image and video segmentation, and MiniMax’s latest open-source agentic model optimized for complex workflows. Models of the week Z.AI’s GLM-4.7-flash Model Basics Model name: zai-org/GLM-4.7-Flash Parameters / size: 30B total -3B active Default settings: 131,072 max new tokens Primary task: Agentic, Reasoning and Coding Why this model matters Why it’s interesting: It utilizes a Mixture-of-Experts (MoE) architecture (30B total parameters and 3B active parameters) to offer a new option for lightweight deployment. It demonstrates strong performance on logic and reasoning benchmarks, outperforming similar sized models like gpt-oss-20b on AIME 25 and GPQA benchmarks. It supports advanced inference features like "Preserved Thinking" mode for multi-turn agentic tasks. Best‑fit use cases: Lightweight local deployment, multi-turn agentic tasks, and logical reasoning applications. What’s notable: From the Foundry catalog, users can deploy on a A100 instance or unsloth/GLM-4.7-Flash-GGUF on a CPU. ource SOTA scores among models of comparable size. Additionally, compared to similarly sized models, GLM-4.7-Flash demonstrates superior frontend and backend development capabilities. Click to see more: https://docs.z.ai Try it Use case Best‑practice prompt pattern Agentic coding (multi‑step repo work, debugging, refactoring) Treat the model as an autonomous coding agent, not a snippet generator. Explicitly require task decomposition and step‑by‑step execution, then a single consolidated result. Long‑context agent workflows (local or low‑cost autonomous agents) Call out long‑horizon consistency and context preservation. Instruct the model to retain earlier assumptions and decisions across turns. Now that you know GLM‑4.7‑Flash works best when you give it a clear goal and let it reason through a bounded task, here’s an example prompt that a product or engineering team might use to identify risks and propose mitigations: You are a software reliability analyst for a mid‑scale SaaS platform. Review recent incident reports, production logs, and customer issues to uncover edge‑case failures outside normal usage (e.g., rare inputs, boundary conditions, timing/concurrency issues, config drift, or unexpected feature interactions). Prioritize low‑frequency, high‑impact risks that standard testing misses. Recommend minimal, low‑cost fixes (validation, guardrails, fallback logic, or documentation). Deliver a concise executive summary with sections: Observed Edge Cases, Root Causes, User Impact, Recommended Lightweight Fixes, and Validation Steps. Meta's Segment Anything 3 (SAM3) Model Basics Model name: facebook/sam3 Parameters / size: 0.9B Primary task: Mask Generation, Promptable Concept Segmentation (PCS) Why this model matters Why it’s interesting: It handles a vastly larger set of open-vocabulary prompts than SAM 2, and unifies image and video segmentation capabilities. It includes a "SAM 3 Tracker" mode that acts as a drop-in replacement for SAM 2 workflows with improved performance. Best‑fit use cases: Open-vocabulary object detection, video object tracking, and automatic mask generation What’s notable: Introduces Promptable Concept Segmentation (PCS), allowing users to find all matching objects (e.g., "dial") via text prompt rather than just single instances. Try it This model enables users to identify specific objects within video footage and isolate them over extended periods. With just one line of code, it is possible to detect multiple similar objects simultaneously. The accompanying GIF demonstrates how SAM3 efficiently highlights players wearing white on the field as they appear and disappear from view. Additional examples are available at the following repository: https://github.com/facebookresearch/sam3/blob/main/assets/player.gif Use case Best‑practice prompt pattern Agentic coding (multi‑step repo work, debugging, refactoring) Treat SAM 3 as a concept detector, not an interactive click tool. Use short, concrete noun‑phrase concept prompts instead of describing the scene or asking questions. Example prompt: “yellow school bus” or “shipping containers”. Avoid verbs or full sentences. Video segmentation + object tracking Specify the same concept prompt once, then apply it across the video sequence. Do not restate the prompt per frame. Let the model maintain identity continuity. Example: “person wearing a red jersey”. Hard‑to‑name or visually subtle objects Use exemplar‑based prompts (image region or box) when text alone is ambiguous. Optionally combine positive and negative exemplars to refine the concept. Avoid over‑constraining with long descriptions. Using the GIF above as a leading example, here is a prompt that shows how SAM 3 turns raw sports footage into structured, reusable data. By identifying and tracking players based on visual concepts like jersey color so that sports leagues can turn tracked data into interactive experiences where automated player identification can relay stats, fun facts, etc when built into a larger application. Here is a prompt that will allow you to start identifying specific players across video: Act as a sports analytics operator analyzing football match footage. Segment and track all football players wearing blue jerseys across the video. Generate pixel‑accurate segmentation masks for each player and assign persistent instance IDs that remain stable during camera movement, zoom, and player occlusion. Exclude referees, opposing team jerseys, sidelines, and crowd. Output frame‑level masks and tracking metadata suitable for overlays, player statistics, and downstream analytics pipelines. MiniMax AI's MiniMax-M2.1 Model Basics Model name: MiniMaxAI/MiniMax-M2.1 Parameters / size: 229B-10B Active Default settings: 200,000 max new tokens Primary task: Agentic and Coding Why this model matters Why it’s interesting: It is optimized for robustness in coding, tool use, and long-horizon planning, outperforming Claude Sonnet 4.5 in multilingual scenarios. It excels in full-stack application development, capable of architecting apps "from zero to one”. Previous coding models focused on Python optimization, M2.1 brings enhanced capabilities in Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript, JavaScript, and other languages. The model delivers exceptional stability across various coding agent frameworks. Best‑fit use cases: Lightweight local deployment, multi-turn agentic tasks, and logical reasoning applications. What’s notable: The release of open-source weights for M2.1 delivers a massive leap over M2 on software engineering leaderboards. https://www.minimax.io/ Try it Use case Best‑practice prompt pattern End‑to‑end agentic coding (multi‑file edits, run‑fix loops) Treat the model as an autonomous coding agent, not a snippet generator. Explicitly require task decomposition and step‑by‑step execution, then a single consolidated result. Long‑horizon tool‑using agents (shell, browser, Python) Explicitly request stepwise planning and sequential tool use. M2.1’s interleaved thinking and improved instruction‑constraint handling are designed for complex, multi‑step analytical tasks that require evidence tracking and coherent synthesis, not conversational back‑and‑forth. Long‑context reasoning & analysis (large documents / logs) Declare the scope and desired output structure up front. MiniMax‑M2.1 performs best when the objective and final artifact are clear, allowing it to manage long context and maintain coherence. Because MiniMax‑M2.1 is designed to act as a long‑horizon analytical agent, it shines when you give it a clear end goal and let it work through large volumes of information—here’s a prompt a risk or compliance team could use in practice: You are a financial risk analysis agent. Analyze the following transaction logs and compliance policy documents to identify potential regulatory violations and systemic risk patterns. Plan your approach before executing. Work through the data step by step, referencing evidence where relevant. Deliver a final report with the following sections: Key Risk Patterns Identified, Supporting Evidence, Potential Regulatory Impact, Recommended Mitigations. Your response should be a complete, executive-ready report, not a conversational draft. Getting started You can deploy open‑source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation. Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft Foundry
