Laguna M.1
Poolside · 2026-04-28 · 225.8B parameters
Laguna M.1 is Poolside’s 225.8B-total Mixture-of-Experts model for agentic coding and long-horizon work, first released on April 28, 2026. It was trained from scratch on more than 30T pre-training tokens drawn from web, code, and synthetic sources, then post-trained through imitation learning and reinforcement learning. Its Apache-2.0 weights support optional interleaved thinking; the 262,144-token window extrapolates a context trained through 128K by doubling the global-attention RoPE scale without additional training.