poolside新出的Laguna XS 2.1，33B MoE专为编程智能体设计，SWE-bench 70.9%还支持Mac本地跑，值得编程党试试
Laguna XS 2.1是poolside推出的33B总参数MoE模型，专为agentic coding和长时任务设计，已获SGLang Day-0支持。模型采用混合SWA+全局注意力（40层中3:1比例）与sigmoid门控，支持FP8 KV缓存和262K上下文，可在36GB RAM的Mac上运行。在SWE-bench Verified上达到70.9%，SWE-bench Multilingual相比XS 2提升5.4%。
🎉 Day-0 support for Laguna XS 2.1 from @poolsideai…
🎉 Day-0 support for Laguna XS 2.1 from @poolsideai is now live in SGLang! This is a 33B total params MoE built for agentic coding and long-horizon work on your local machine. 1️⃣ Native interleaved thinking between tool calls, toggle per-request 2️⃣ Mixed SWA + global attention (3:1 across 40 layers) with sigmoid gating 3️⃣ FP8 KV cache + 262K context — runs on a Mac with 36GB RAM 4️⃣ 70.9% SWE-bench Verified, +5.4% jump on SWE-bench Multilingual vs XS.2
Cookbook: https://t.co/kUDo7yV4f6 Run it now with SGLang!