Summary
Poolside AI has released two new open-weight coding models: Laguna M.1 (2–5 billion parameters) and Laguna XS.2, a 33-billion-parameter mixture-of-experts model with only 3 billion parameters activated per token. In this video, Fahd Mirza installs and tests Laguna XS.2 locally on an NVIDIA H100, walking through a notably involved setup that required building vLLM from source — a process taking roughly three to five hours because official vLLM support for the model had not yet landed in the stable release at time of recording.
The model’s architecture includes 40 transformer layers (30 using 512-token sliding window attention, 10 with full global attention and per-head sigmoid gating), FP8 KV cache quantization, native reasoning with interleaved thinking between tool calls, and a 128k context window. It uses 256 experts plus one shared expert and was trained with async off-policy agent RL using the Muon optimizer, targeting long-horizon agentic coding tasks. VRAM consumption at load sits around 76GB on the H100.
Testing shows mixed results: a real-time collaborative code editor prompt produced non-functional output, while a self-contained Kanban board HTML task showed partial success. Mirza benchmarks these outputs against recent testing of DeepSeek V4, providing honest comparative context. The model is released under Apache 2.0 and is also available via Ollama for simpler local deployments, making this video a practical guide for developers evaluating open-weight coding agents.
📺 Source: Fahd Mirza · Published April 29, 2026
🏷️ Format: Hands On Build
