6 hours ago
- Laguna S 2.1 is a 118B-parameter Mixture-of-Experts model with 8B active parameters per token and a 1M-token context window, developed in under nine weeks.
- It achieves competitive scores against much larger models on long-horizon coding benchmarks, scoring 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1.
- All evaluation trajectories are publicly released at trajectories.poolside.ai for transparency and reproducibility.
- Case studies show the model building a browser engine from scratch, optimizing its own harness (5.2% speedup, 70% memory reduction), and rediscovering a mathematical proof.
- The model offers two thinking modes: off and max, with max significantly improving performance (e.g., Terminal-Bench 2.1 from 60.4% to 70.2%).
- Known limitations include harness overfitting, nested tool call issues, and longer than expected thinking duration on difficult problems.
- Key training improvements include larger rollout budgets, better sandbox infrastructure, multi-harness rollouts, and RL in FP8 precision.
- The model focuses on agentic coding capabilities and on recovering hidden reasoning through reinforcement learning.
- Laguna S 2.1 is available via multiple platforms (Hugging Face, OpenRouter, Baseten) with open weights under OpenMDW-1.1.