Poolside ships Laguna S 2.1, a 118B open-weight coding MoE

Poolside published Laguna S 2.1 on Hugging Face, a 118B-parameter Mixture-of-Experts model with roughly 8B parameters activated per token. It sits in the middle of the Laguna line between a 33B smaller sibling (Laguna XS 2.1, 33B-A3B) and a 225B larger one (Laguna M.1, 225B-A23B). The architecture is where the interesting choices are: 256 routed experts with a top-10 router plus one shared expert, 48 layers in a 1:3 ratio of global to sliding-window attention, and a 1,048,576-token context window.

On the benchmark table Poolside itself published, Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench Multilingual, 59.4% on SWE-Bench Pro, 40.4% on DeepSWE and 49.7% on Toolathlon Verified. Against DeepSeek-V4-Pro Max at 1.6T-A49B, those numbers are competitive-to-better on most of the coding evals Poolside chose. But Tencent Hy3 (295B-A21B) edges it on Terminal-Bench at 71.7%, and Poolside's own table shows Kimi K3 at 88.3% on Terminal-Bench with 69% on DeepSWE, and Claude Fable 5 at 88% on Terminal-Bench and 80.3% on SWE-Bench Pro. The honest read is that Laguna S 2.1 is a credible open-weight coder that outperforms DeepSeek on Poolside's chosen evals but is not the top of any leaderboard.

The part that will actually move procurement decisions is the license and the packaging. The weights ship under OpenMDW-1.1, described in the card as permitting free commercial and non-commercial use and modification. That is a meaningfully more permissive posture than research-only or 'community' licenses, and it is the specific unlock for enterprises that want an agentic coding model running inside their own network without a per-token bill. vLLM and SGLang serving recipes ship in the card, along with FP8, NVFP4, INT4 and GGUF quantized variants.

The honest caveat is that at 118B parameters, in Poolside's own words, the BF16 checkpoint needs multiple GPUs, roughly 236GB of weights, so 'open weights' is not the same as cheap self-hosting for a small team. Coding benchmarks are leaky, and the model card doesn't publish long-context evals, which matters because the 1M window is one of the marquee specs. The forward-looking question is whether OpenMDW-1.1 becomes a template other labs adopt: if it does, the balance between hosted-only frontier coders and self-hostable ones shifts faster than the raw benchmark numbers alone would suggest.