Background

With the successful merge of PR for MindSpore support, we are issuing the roadmap for 2025 Q4. This is a follow-up of our previous roadmap for 2025 H2.

MindSpore Model Repo

Currently, all MindSpore models reside in a separate repo and needs to be installed separately: https://github.com/chz34/sgl-mindspore. We plan to move the repo under sgl-project if it proves to be a smooth user experience.

SGLang Features

• [November] PD Disaggregation - initial experiments already worked for Qwen3 models.
• [December] EP on DeepSeek V3 and Qwen3 MoE
• [December] EAGLE-3 Speculative Decoding for Qwen3 dense and MoE
• [December] W8A8 Quantization
• [December] Multi-LoRA for Qwen3

CI tests

• [December] e2e/unit test coverage

User / Developer Experience

• Up-to-date Dockerfile and docker image
• Up-to-date Documentations (installation, quickstart, tutorials, etc.)

Community

Actively use the #sglang-mindspore slack channel as a platform for discussion.

Comments and suggestions are welcome!