Poolside releases the weights for Laguna M.1, its flagship code model

Poolside releases open weights for Laguna M.1, its 225.8B MoE coding model with a 256K context window, now available on Hugging Face under Apache 2.0.

Poolside is releasing the weights for Laguna M.1, which it describes as its most capable model to date, now in open weights. Both base and post-trained versions are available on Hugging Face under an Apache 2.0 license, with a 256K context window.

M.1 is a Mixture-of-Experts (MoE) built for long-context agentic coding, boasting 225.8 billion total parameters with 23.4 billion activated per token, and trained from scratch on 6,144 NVIDIA H200 GPUs. Previously available via Poolside's API and OpenRouter, including by code agents like Kilo Code and Nous Research's Hermes agent, it can now run on anyone's infrastructure, be evaluated in their own harnesses, and fine-tuned.

Poolside is making open weights its new standard. M.1 and its smaller sibling XS.2, already open and compact enough for a single GPU, remain free for now on the API and OpenRouter, with dedicated paid endpoints coming soon for intensive use. In terms of performance, the company positions M.1 as competitive with the best open and frontier models on agentic coding benchmarks, and reports around 72.5% on SWE-bench Verified. It encourages users to use it via pool, its agent harness that acts as both an ACP server and client, which is the environment it uses internally.