Introducing Laguna S 2.1

Laguna S 2.1 is Poolside's 118B-total-parameter Mixture-of-Experts coding model (8B active per token), released July 2026 with a 1M-token context window. Trained in under nine weeks on 4,096 NVIDIA H200 GPUs, it scores 70.2% on Terminal-Bench 2.1 with thinking enabled — outperforming much larger models like DeepSeek-V4-Pro-Max (1.6T params, 64.0%) and Nemotron 3 Ultra (550B, 56.4%).

The model supports two thinking modes (off and max), with max thinking lifting DeepSWE scores from 16.5% to 40.4%. It runs on a single NVIDIA DGX Spark, targeting local-machine agentic coding as its niche. RL post-training was done in FP8 precision — a first for Poolside — across 409k agentic and non-agentic environments, with new multi-harness rollouts to prevent overfitting.

Three case studies demonstrate capability: building a full HTML/CSS rendering engine from scratch in a single 50-minute session, optimizing Poolside's own agent harness for a 5.2% speedup, and independently rediscovering a proof to Erdős problem #397 using only Perl for brute-force work. Acknowledged limitations include harness overfitting (relying on memory of tool interfaces rather than reading schemas), nested-tool-call JSON errors, and occasional overthinking on competition math.

The release is available on Hugging Face under OpenMDW-1.1, with day-one inference support in vLLM, SGLang, Ollama, and TRT-LLM, plus hosted access via OpenRouter and Baseten.