v4.5.0

New Models

Qwen3.8

Qwen3.8, the strongest open Qwen generation so far. Same Gated DeltaNet hybrid backbone as Qwen3.5/3.6, native vision/video, hybrid thinking on by default, plus tunable reasoning effort and preserved historical thinking. (#9914)

NVIDIA Nemotron 3.5 Lightning

Nemotron 3.5 Lightning 30B-A3B: hybrid Mamba-2 + MoE + Attention, 30B total / 3B active, 1M context, MTP. (#9894)

Muse Glimmer

Muse-Glimmer-30B, Meta’s 30B dense multimodal model distilled from Muse Spark for local agents. (#9887)

Other changes

• MiMo-V2.5 inference (vLLM / SGLang) (#9880)
• DeepSeek-V4-Flash-0731 adaptation (#9834)
• Thinking-prefix loss scale (#9875)
• OpenAI / Anthropic message processors (#9809)

---

What's Changed

• fix: preserve explicit PEFT LoRA dtype by @yaodong-shen in #9832
• Fix export merge_lora with quantization not working by @hyongtao-code in #9830
• Support OpenAI and Anthropic message processors by @lumincui in #9809
• [bugfix] Fix Qwen3.5 non-packing causal conv routing by @taking-lying-flat in #9812
• Fix Gemma 3 mixed-batch token type IDs by @GoGiants1 in #9786
• remove qoder-review action by @hjh0119 in #9842
• Fix GKD Liger teacher routing by @taking-lying-flat in #9841
• fix(infer): propagate streaming errors by @RerankerGuo in #9849
• fix(npu): handle BF16 MindSpeed grouped linear activation offload by @addsubmuldiv in #9846
• Fix GRPO offload compatibility with FSDP2 CPU offload by @addsubmuldiv in #9844
• fix(infer): reject prompts without generation space by @RerankerGuo in #9851
• docs(readme): fix v3.0 release note links by @RerankerGuo in #9853
• Support ds v4 flash 0731 by @tastelikefeet in #9834
• Fix missing LoRARequest injection in GRPOVllmEngine.infer_async by @Zuozhuo in #9856
• fix(utils): close daemon event loops on shutdown by @RerankerGuo in #9858
• fix(utils): preserve streamed download bytes by @RerankerGuo in #9857
• fix(qwen3.5): add Torch fallback for Qwen3.5 linear attention SP by @meichangsu1 in #9865
• update 4.4.3 image by @hjh0119 in #9869
• Support empty content openai message format and standalone tool message by @tastelikefeet in #9861
• feat: add loss mask for response-start thinking prefixes by @yph22 in #9875
• [Infer] Fix transformers streaming logprobs by @taking-lying-flat in #9866
• support Xiaomi-MiMo-V2.5 inference( sglang / vllm ) by @z0o0ey in #9880
• fix(template): keep MiMo dependencies optional by @RerankerGuo in #9886
• Support muse-glimmer by @tastelikefeet in #9887
• fix(rlhf): reduce sequence-parallel log-prob memory by @yongyaoduan in #9882
• [model] support nemotron by @hjh0119 in #9894
• Fix MathORM false positives on parse failures by @Ruihan11 in #9889
• Fix dcp_validation bug in megatron distributed saving by @tastelikefeet in #9896
• fix:MiMo audio input by @z0o0ey in #9898
• Fix sp with transformers5.15 by @tastelikefeet in #9899
• Synchronize MuonClip max logits across ranks by @taking-lying-flat in #9901
• fix: add set_epoch() to DataLoaderDispatcher for streaming datasets by @tomzw11 in #9902
• Fix alpha UMI Next loss scale by @taking-lying-flat in #9900
• fix(megatron/grpo): clear MegatronGRPOTrainer._metrics per rollout by @HSYZhang in #9884
• fix megatron common pt save by @hjh0119 in #9908
• fix nemotron grpo by @hjh0119 in #9905
• remove auto_config patch for vllm by @hjh0119 in #9911
• [model] qwen3.8 by @hjh0119 in #9914
• rename qwen3.5 doc by @hjh0119 in #9918
• fix(README): restore broken star history chart by @PingouinFerreux in #9913

New Contributors

• @yaodong-shen made their first contribution in #9832
• @lumincui made their first contribution in #9809
• @GoGiants1 made their first contribution in #9786
• @RerankerGuo made their first contribution in #9849
• @Zuozhuo made their first contribution in #9856
• @yph22 made their first contribution in #9875
• @yongyaoduan made their first contribution in #9882
• @Ruihan11 made their first contribution in #9889
• @tomzw11 made their first contribution in #9902
• @HSYZhang made their first contribution in #9884
• @PingouinFerreux made their first contribution in #9913

Full Changelog: v4.4.3...v4.5.0