Top AI Product

Pre-training started May 22 on 4,096 H200s. The model shipped July 21. That’s under nine weeks from zero to a coding model that beats Qwen 3.7 Max on SWE-Bench Multilingual.

Laguna S 2.1 is an open-weight MoE model built for long-horizon agentic coding — 118B total parameters, 8B active per token, 1M context. Not a chat assistant. It’s the thing you hand a blank folder and 50 minutes, and it runs 181 steps to build an HTML/CSS rendering engine. Thinking mode is on by default and it pays for itself: Terminal-Bench 2.1 jumps from 60.4% to 70.2%. SWE-Bench Multilingual hits 78.5%, SWE-Bench Pro 59.4%.

How to run it

Weights are on Hugging Face under OpenMDW-1.1 in BF16, FP8, INT4 and NVFP4, plus official GGUF and MLX builds. OpenRouter serves it at $0.10 in / $0.20 out per million tokens, with a free 256K endpoint and paid 1M context. Baseten, Vercel AI Gateway, vLLM and Ollama all host it. Cline, OpenCode and Hermes Agent already ship it as a backend. chat.poolside.ai needs no login.

Why it matters

8B active params means it fits on a single DGX Spark — a frontier-ish agentic coder you can leave running overnight without watching a meter. That’s the whole argument. Poolside burned three years and roughly $600M to get here, and HN put it at 259 points in a day. The pitch is blunt: the West’s answer to DeepSeek and Qwen, at open weights and pennies per million tokens.

You Might Also Like

• Z ais glm 5 1 Tops swe Bench pro at 58 4 Trained on Zero Nvidia Hardware
• Mistral Medium 3 5 Scores 77 6 on swe Bench one Point shy of Gemini 3 1 pro
• Kimi k2 6 Beats gpt 5 4 and Claude Opus 4 6 on swe Bench pro
• Ornith 1 0 Deepreinforce Self Scaffolding Coding Models Open Weights That Write Their own rl Scaffold 397b Hits 82 4 on swe Bench
• Claude Sonnet 5 Scores 63 2 on swe Bench pro at a Third of Opus 4 8s Price