NVFP4 Quantized RedHatAI/Qwen3.6-35B-A3B-NVFP4
This is a preliminary version (and subject to change) of NVFP4 quantized Qwen/Qwen3.6-35B-A3B model. The model has both weights and activations quantized to NVFP4 format with vllm-project/llm-compressor.
It is compatible and tested against vllm main. Deploy it with: vllm serve RedHatAI/Qwen3.6-35B-A3B-NVFP4 --reasoning-parser qwen3 --moe_backend flashinfer_cutlass.
If you have hardware with more compute than memory bandwidth, you may prefer this MoE variant for performance reasons.
Creation Script:
Run this script with LLM Compressor main and latest transformers.
import torch
from compressed_tensors.utils import save_mtp_tensors_to_checkpoint
from datasets import load_dataset
from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration
from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
# NOTE: This example requires transformers >= v5
MODEL_ID = "Qwen/Qwen3.6-35B-A3B"
# Load model.
model = Qwen3_5MoeForConditionalGeneration.from_pretrained(MODEL_ID, dtype="auto")
processor = AutoProcessor.from_pretrained(MODEL_ID)
# No need to include mtp layers as they are not loaded
# through Qwen3_5MoeForConditionalGeneration
recipe = QuantizationModifier(
targets="Linear",
scheme="NVFP4",
ignore=[
"re:.*lm_head",
"re:visual.*",
"re:model.visual.*",
"re:.*mlp.gate$",
"re:.*embed_tokens$",
"re:.*shared_expert_gate$",
"re:.*linear_attn.*",
],
)
NUM_CALIBRATION_SAMPLES = 256
MAX_SEQUENCE_LENGTH = 4096
ds = load_dataset(
"HuggingFaceH4/ultrachat_200k",
split=f"train_sft[:{NUM_CALIBRATION_SAMPLES}]",
)
ds = ds.select_columns(["messages"])
ds = ds.shuffle(seed=42)
def preprocess_function(example):
messages = [
{"role": m["role"], "content": [{"type": "text", "text": m["content"]}]}
for m in example["messages"]
]
return processor.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
add_generation_prompt=False,
processor_kwargs={
"return_tensors": "pt",
"padding": False,
"truncation": True,
"max_length": MAX_SEQUENCE_LENGTH,
"add_special_tokens": False,
},
)
ds = ds.map(preprocess_function, batched=False, remove_columns=ds.column_names)
def data_collator(batch):
assert len(batch) == 1
return {key: torch.tensor(value) for key, value in batch[0].items()}
# Apply quantization.
oneshot(
model=model,
recipe=recipe,
dataset=ds,
max_seq_length=MAX_SEQUENCE_LENGTH,
num_calibration_samples=NUM_CALIBRATION_SAMPLES,
moe_calibrate_all_experts=True,
data_collator=data_collator,
)
# Save to disk in compressed-tensors format.
SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-NVFP4"
model.save_pretrained(SAVE_DIR)
processor.save_pretrained(SAVE_DIR)
# MTP layers are excluded from the model through Qwen3_5MoeForConditionalGeneration
# Save them as-is from the original checkpoint into the quantized output.
save_mtp_tensors_to_checkpoint(source_model=MODEL_ID, dest_dir=SAVE_DIR)
Evaluation
This model was evaluated on GSM8K-Platinum, MMLU-Pro, IFEval, Math 500, GPQA Diamond, AIME 25, and LiveCodeBench v6 using lm-evaluation-harness and lighteval, served with vLLM using --language-model-only.
Accuracy
| Benchmark | Qwen/Qwen3.6-35B-A3B | RedHatAI/Qwen3.6-35B-A3B-NVFP4 | Recovery (%) |
|---|---|---|---|
| GSM8k Platinum (0-shot) | 95.73 | 96.08 | 100.37 |
| IfEval (0-shot) | 93.09 | 92.45 | 99.31 |
| AIME 2025 | 92.92 | 91.25 | 98.21 |
| GPQA diamond | 84.51 | 84.68 | 100.20 |
| Math 500 | 84.80 | 85.00 | 100.24 |
| Lcb Codegeneration V6 | 77.33 | 74.67 | 96.55 |
| MMLU Pro Chat | 85.32 | 84.70 | 99.28 |
| BFCLv4 Overall | 57.83 | 56.10 | 97.01% |
| BFCLv4 Single Turn | 53.81 | 53.45 | 99.34% |
| BFCLv4 Multi-Turn | 62.25 | 58.13 | 93.38% |
| BFCLv4 Agentic | 49.91 | 49.31 | 98.80% |
Reproduction
The results were obtained using the following commands:
The model was served with vLLM using the following command:
vllm serve RedHatAI/Qwen3.6-35B-A3B-NVFP4 --reasoning-parser qwen3 --language-model-only --max-model-len 96000
Each benchmark was run 3 times with different seeds (42, 1234, 4158), except AIME 25 which used 8 seeds (42, 1234, 4158, 5322, 1356, 9843, 3344, 5678). Scores are averaged across all seeds.
lm-eval benchmarks
GSM8K-Platinum (0-shot)
lm_eval --model local-chat-completions \
--tasks gsm8k_platinum_cot_llama \
--model_args "model=RedHatAI/Qwen3.6-35B-A3B-NVFP4,max_length=96000,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=128,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=2400" \
--num_fewshot 0 \
--apply_chat_template \
--output_path results.json \
--seed 42 \
--gen_kwargs "do_sample=true,temperature=1.0,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000,presence_penalty=1.5,repetition_penalty=1.0,seed=42"
IFEval (0-shot)
lm_eval --model local-chat-completions \
--tasks ifeval \
--model_args "model=RedHatAI/Qwen3.6-35B-A3B-NVFP4,max_length=96000,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=128,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=2400" \
--apply_chat_template \
--output_path results.json \
--seed 42 \
--gen_kwargs "do_sample=true,temperature=1.0,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000,presence_penalty=1.5,repetition_penalty=1.0,seed=42"
MMLU-Pro (0-shot)
lm_eval --model local-chat-completions \
--tasks mmlu_pro_chat \
--model_args "model=RedHatAI/Qwen3.6-35B-A3B-NVFP4,max_length=96000,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=128,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=3600" \
--num_fewshot 0 \
--apply_chat_template \
--output_path results.json \
--seed 42 \
--gen_kwargs "do_sample=true,temperature=1.0,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000,presence_penalty=1.5,repetition_penalty=1.0,seed=42"
lighteval benchmarks
litellm_config.yaml:
model_parameters:
provider: "hosted_vllm"
model_name: "hosted_vllm/RedHatAI/Qwen3.6-35B-A3B-NVFP4"
base_url: "http://0.0.0.0:8000/v1"
api_key: ""
timeout: 2400
concurrent_requests: 64
generation_parameters:
temperature: 1.0
max_new_tokens: 64000
top_p: 0.95
top_k: 20
min_p: 0.0
presence_penalty: 1.5
repetition_penalty: 1.0
seed: 0
Math 500, GPQA Diamond, LiveCodeBench v6 (0-shot)
lighteval endpoint litellm litellm_config.yaml \
"math_500|0,gpqa:diamond|0,lcb:codegeneration_v6|0" \
--output-dir results \
--save-details
AIME 25 (0-shot)
lighteval endpoint litellm litellm_config.yaml \
"aime25|0" \
--output-dir results \
--save-details
BFCLv4
BFCL requires the model to be registered in the leaderboard codebase before running evaluation.
Step 1 — Register the model in bfcl_eval/constants/model_config.py
Add the following entry to api_inference_model_map:
"Qwen3.6-35B-A3B-NVFP4": ModelConfig(
model_name="Qwen3.6-35B-A3B-NVFP4",
display_name="Qwen3.6-35B-A3B-NVFP4 (FC)",
url="https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4",
org="Google",
license="Apache 2.0",
model_handler=OpenAICompletionsHandler,
input_price=None,
output_price=None,
is_fc_model=True,
underscore_to_dot=True,
),
Step 2 — Add the key to bfcl_eval/constants/supported_models.py
Add "Qwen3.6-35B-A3B-NVFP4" to the SUPPORTED_MODELS list.
Step 3 — Start the vLLM server (use the command at the top of this section; the --served-model-name flag ensures BFCL can find the model by its registered slug).
Step 4 — Generate responses and evaluate
bfcl generate --model Qwen3.6-35B-A3B-NVFP4 --test-category all
bfcl evaluate --model Qwen3.6-35B-A3B-NVFP4 --test-category all
- Downloads last month
- 1,959,618