Laguna M.1

Poolside's flagship agentic-coding model: a 225.8B-total / 23.4B-active Mixture-of-Experts transformer (256 experts, top-8 plus one shared expert; first layer dense; pre-norm RMSNorm) with a 256K-token context. Trained from scratch in-house on >30T tokens using 6,144 NVIDIA H200 GPUs with the Muon (Moonlight-variant) optimizer, then post-trained via mid-training, SFT, and an asynchronous online-RL stack using a CISPO objective on verifiable software-engineering rewards (the team compared CISPO against GRPO and GSPO and chose it).

Originally launched April 2026 as an API-only model, Laguna M.1 was released as open weights under Apache 2.0 in June 2026 — both base and post-trained checkpoints, plus FP8 and NVFP4 quantizations, on Hugging Face — as Poolside made open releases its default. Serves on vLLM and SGLang; full BF16 inference requires multi-GPU memory.

Reported benchmarks (technical report, May 2026): SWE-bench Verified 74.6, SWE-bench Multilingual 73.3, SWE-bench Pro 49.2, Terminal-Bench 2.0 56.9 — frontier-tier on agentic coding, ahead of the Devstral 2, GLM-4.7, DeepSeek-V4-Flash, Qwen3.5, and Claude Sonnet 4.6 comparators in the report's SWE-bench Verified figure. Not currently scored on Artificial Analysis — numbers above are self-reported.