Poolside releases Laguna S 2.1, the West's most capable open-weight model

What happened

Poolside released Laguna S 2.1 on July 21, 2026 — a 118B Mixture-of-Experts model with 8B active parameters per token and a 1 million token context window. Training completed in 9 weeks from May 22, 2026. Weights are freely available on Hugging Face under the OpenMDW-1.1 license and the model runs on a single NVIDIA DGX Spark.

Context and impact

Poolside calls it 'the most capable agentic coding model in its weight class by a wide margin.' The model beats DeepSeek-V4-Flash, NVIDIA Nemotron 3 Ultra, and Thinking Machines Inkling despite being 10x smaller — demonstrating the power of RL from code execution feedback.

Details

• Architecture: 118B MoE, 8B active params/token; 1M token context window
• Benchmarks: DeepSWE v1.1: 40.4% (+24 pp with thinking mode); SWE-Bench Pro: 59.4%; SWE-Bench Multilingual: 78.5%
• License: OpenMDW-1.1 — free on Hugging Face
• Access: chat.poolside.ai, OpenRouter ($0.10/$0.20 per MTok), vLLM, SGLang, Ollama
• Training: from scratch on own infrastructure via RL from code execution feedback