sgl-project/sglang v0.5.22
| Model | Type | Cookbook |
ProofCode ↗
Key points
- | LLaDA2.2 | LLM / VLM | |
- | LLaDA-Image / LLaDA-Image-Turbo | Diffusion | |
- Important fix: Hybrid linear models could resume from wrong mamba states when speculative decoding and structured output were used together.
Sources (1)
- [1]sgl-project/sglang v0.5.22GitHub: sgl-project/sglang · Oct 19, 02:55 AM
| Model | Type | Cookbook |
| LLaDA2.2 | LLM / VLM | |
Extractive summary: sentences quoted from the sources.
Before this
- Oct 18, 2026vllm-project/vllm v0.32.0
- Oct 17, 2026Nemotron 3.5 Bolt, Muse Spark Open, and a Fast LFM2.5 24B
- Oct 16, 2026Welcome Beam: a 501B open model you can own, customize, and scale
- Oct 15, 2026unslothai/unsloth v0.1.908-beta: NVIDIA Nemotron 3.5 + Meta Muse Spark Open
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- Oct 15, 2026StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context
