Poolside Drops Laguna M.1, a 225B Open-Weight Coding Agent

Poolside has released Laguna M.1, an Apache 2.0-licensed coding model built for agentic software engineering. Its Mixture-of-Experts architecture contains 225 billion parameters but activates about 23 billion for each token, reducing computation while still requiring the full set of weights in memory.

• Context window: 262,144 tokens
• License: Apache 2.0
• Deployment: Self-hosted weights or Poolside’s hosted API
• Primary workload: Multi-step coding tasks with shell and file-editing tools

The release gives developers a self-hostable alternative for repository-scale coding agents, including environments where source code cannot be sent to a closed model provider. Poolside also published quantized checkpoints, integrations with major inference frameworks, and details about the training system behind the model.

Before Laguna M.1, Poolside had concentrated on government and public-sector deployments, including systems designed for high-security environments. The company, co-founded by former GitHub CTO Jason Warner, describes Laguna as its first open-weights release.

256 experts, 23B active parameters

Laguna M.1 uses a 70-layer transformer. Its first three layers are dense SwiGLU blocks, while the remaining 67 are sparse MoE layers containing 256 routed experts and one shared expert. For each token, the router selects 16 specialized experts in addition to the shared expert, a wider fan-out than the two to eight experts common in many MoE models.

• Global attention: Every layer can attend across the full prompt, without a sliding window.
• Grouped-query attention: The model has 64 query heads and eight key-value heads, each with a dimension of 128. Fewer key-value heads reduce cache requirements.
• Attention gating: A learned Softplus gate scales each attention head’s output.
• Long-context encoding: Rotary position embeddings use YaRN scaling to reach 262,144 tokens.
• Router balancing: Expert utilization is balanced without an auxiliary routing loss.

Global attention gives every layer access to the full sequence, but long prompts still increase prefill time, memory use, and attention cost. The advertised context limit therefore describes capacity rather than a fixed operating target for every deployment.