Laguna S 2.1
The West's open-weights answer, licensed under Linux Foundation paper
What it is
Poolside's 118B-total / 8B-active MoE coding model with a 1M-token context and switchable thinking modes — pre-trained in under nine weeks on 4,096 H200s. Ships in BF16 plus official FP8, INT4, and NVFP4 builds, an XS sibling, and community GGUFs; a free no-login demo runs at chat.poolside.ai.
Why it's interesting
70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE puts it in the ring with models 10–20× its active size, and it's the rare Western frontier-lab release with weights you can pull today. The OpenMDW-1.1 license is the story: Linux Foundation-backed, MIT-grade in substance — and the same license we already had to make a judgment call on for Cosmos 3.
Use cases
- Self-hosted agentic coding backends with 1M-token context
- Terminal and repo-scale automation on a single DGX-class box
- Fine-tuning a competitive Western base for code domains
Who it's for
Coding-agent builders; teams wanting non-PRC open weights
Setup
Moderate. Runs on a single NVIDIA DGX Spark; official INT4/NVFP4 and community GGUF builds bring it toward workstation reach
Limitations & cautions
Benchmarks concentrate on coding and terminal work rather than general chat, several hosts cap context below 1M, and independent evaluation of the launch numbers is still thin.
Editorial takeaway
poolside spent years closed. Its first open drop uses the most open model license on the shelf — the only thing between it and 'open' is OSI's docket.