Development Roadmap (2025 Q4)

[sii-xinglong]

Here is the development roadmap for 2025 Q4. Contributions and feedback are welcome(Open a new discussion)

Focus

• More model support
• Improve Language models performance
• Reinforcement learning training framework integration.

Model coverage

Language models

• Grok2
• Ling 2.0
• Qwen3-235B-A22B
• GPT-OSS [Feature] Support gpt-oss-20b model #407
• DeepSeek Series [Feature] Support DeepSeek Series Models #420

Multi-modal models

• MiMo-Audio
• Wan2.1
• Qwen2.5-VL [Feature] support qwen 2.5vl #225

Speculative decoding

• Reference-based speculative decoding support
• Feat/eagle support #378
• Feat/refactor draft ar rebase #544
• [Feat][Eagle]refactor verify phase rebase #616

RL framework integration

• sglang-jax serves as the inference engine backend for tunix
  • Accuracy check: Test math-500 and aime24 for deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
    • [docs] add evaluations math #286
  • Support return_logprobs: return logprobs as float32
    • add return logprobs #310
  • Support pathways mode: single host & multi host
    • [Feature] Support pathways mode #326
  • Interruptible Sampling [Feature-RL] Support Interruptible Sampling For Tunix From Rollout Angle #447

Performance improvements

• Continue to optimize the implemented model and use the GPU version as the baseline

LoRA Support

• Support multi LoRA servering
  Issue: [Feature] Add Multi LoRA Support #311

Deterministic inference

• [Feature] Support deterministic inference #325

Data Parallelism

• [Feature] Data-Parallel(DP) Support #3