SGLang supported models: Large Language Models (generative)

Read 2026-10-03. Confirms the hybrid/SSM/SWA families BY MODEL NAME rather than by architecture class, with the mechanism stated in prose: Granite 4.0 micro described as 'hybrid Mamba-MoE ... hybrid language model combining gated short convolutions with a small number of grouped query attention (GQA) blocks'; Kimi Linear 'hybrid linear attention model (48B total, 3B active) with 1M token context; features Kimi Delta Attention (KDA) for up to 6x faster decoding and 75% KV cache reduction vs full attention'; Nemotron Nano 'hybrid Mamba-Transformer architecture combining attention and state-space models'; Ling described as 'hybrid MoE model: per-layer mix of DeepSeek-style MLA full attention and Qwen3 Gated-Delta-Net (GDN) linear attention, DeepSeek-style routed + shared experts, and multiple multi-token-prediction (MTP) heads'; GPT-OSS described as 'hybrid sliding-window-attention MoE model (256 routed experts, 1 shared expert, sigmoid router) with per-layer-type RoPE (YARN on full-attention layers, default on sliding-attention layers) and a softplus per-head attention gate'; LFM2 'hybrid backbone, with SwiGLU experts, normalized sigmoid routing, and top-k expert selection'; Mamba-Codestral 'Mamba2 state-space model for code generation; uses selective state spaces instead of attention, so it holds no KV cache'. Also the page's own caveat on how to check coverage: search the sglang repo for the architecture class string. Qwen3.5-397B-A17B is listed under generative, Qwen3-Next-80B-A3B-Instruct listed, Falcon-H1-34B-Instruct listed, granite-swash SWA models listed.