Poolside Releases Laguna M.1: 225B Parameter MoE for Agentic Coding

Poolside has released Laguna M.1, a 225B total parameter Mixture-of-Experts (MoE) model with 23B activated parameters per token, designed specifically for agentic coding and long-horizon development work.

Architecture and Specifications

Laguna M.1 uses a 70-layer MoE transformer with 256 experts and top-k=16 routing. The first 3 layers are dense SwiGLU, while the remaining 67 layers use sparse MoE. The model employs global attention across all layers with 64 Q-heads, 8 KV-heads, and head dimension 128. Context window extends to 262,144 tokens.

The model uses RoPE with YaRN for positional encoding and includes native reasoning support through interleaved thinking between tool calls. Training involved pre-training, post-training, and reinforcement learning stages using the Muon optimizer.

Benchmark Performance

Poolside reports the following scores on agentic coding benchmarks (averaged over 4 runs at temperature=1.0, top_k=20):

• SWE-bench Verified: 74.6%
• SWE-bench Multilingual: 63.1%
• SWE-bench Pro: 49.2%
• Terminal-Bench 2.0: 45.8%

For comparison, DeepSeek-V4 Flash (284B total, 13B active) scores 79.0% on SWE-bench Verified, while Qwen3.5 (397B total, 17B active) achieves 76.2%. Claude Sonnet 4.6 scored 79.6% on SWE-bench Verified and 59.1% on Terminal-Bench 2.0, according to Poolside's comparison table.

Deployment and Availability

The model is released under Apache 2.0 license and supports deployment via vLLM (version 0.21.0+), SGLang, Transformers (v5.7.0+), and TensorRT-LLM. Quantized checkpoints are available in FP8 and NVFP4 formats.

Poolside provides a terminal-based coding agent called "pool" that integrates with the model and supports the Agent Client Protocol. The tool auto-configures with Zed and JetBrains editors.

All benchmarking used Poolside's pool agent harness with maximum 500 steps and sandboxed execution. Tasks ran in 8GB RAM/2 CPU sandboxes, except Terminal-Bench 2.0 which used 48GB RAM/32 CPUs. Poolside states they ran a reward-hack judge post-hoc and "did not find significant reward hacking."

What This Means

Laguna M.1 represents a competitive open-weight option for agentic coding tasks, though it trails frontier models like DeepSeek-V4 Flash on SWE-bench benchmarks by 4-11 percentage points depending on the task. The Apache 2.0 license and support for standard serving frameworks lower deployment barriers compared to proprietary alternatives. The 225B total parameter count with 23B active parameters positions it between smaller dense models and larger MoE architectures in terms of inference cost, though actual performance-per-dollar will depend on hardware and optimization.