poolside刚发的225B MoE模型Laguna M.1,专为智能体编码设计,SGLang直接跑起来了,在SWE-bench上很强。
poolside发布的Laguna M.1是一个225B参数的MoE模型,专为智能体编码和长期任务设计。该模型采用70层结构:3个密集SwiGLU层加67个稀疏MoE层,共有256个专家,top-k=16且使用无辅助损失负载均衡。它在所有层使用全局注意力:64个Q头、8个KV头,以及softplus输出门控。Laguna M.1支持原生交错推理:在工具调用之间进行思考,并可每个请求切换。在SWE-bench Verified、SWE-bench Multilingual、SWE-Bench Pro和Terminal-Bench 2.0上表现强劲。现在可通过SGLang运行。
🎉 Day-0 support for Laguna M.1 from @poolsideai is…
🎉 Day-0 support for Laguna M.1 from @poolsideai is live in SGLang! This is a 225B MoE built for agentic coding & long-horizon work. 1️⃣ 70-layer MoE: 3 dense SwiGLU layers + 67 sparse MoE layers, 256 experts, top-k=16 with aux-loss-free load balancing 2️⃣ Global attention across all layers: 64 Q-heads, 8 KV-heads, softplus output gating 3️⃣ Native interleaved reasoning: thinking between tool calls, toggleable per-request 4️⃣ Strong on SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro & Terminal-Bench 2.0.
Run it now with SGLang!