Poolside Just Released Laguna S 2.1 — And It's a Big Deal for Open-Weight AI

San Francisco-based AI startup Poolside has released Laguna S 2.1, a 118-billion-parameter open-weight model built for agentic coding. It's already being talked about as one of the most important open-weight releases of the year — not because it's the single best coding model on the planet, but because of who released it, how small it is, and what it can do with that size.

Laguna S 2.1 is a mixture-of-experts model with 8B activated parameters per token and supports a context window of up to 1M tokens in thinking and no-thinking modes. On OpenRouter, it's currently accessible for free at a 262K context window, though the full model supports far more. It sits in the middle of a three-model family, between a smaller sibling and a larger one released by Poolside.

This isn't just another model drop — it's being framed explicitly as a response to a specific gap in the market. As Poolside describes it, over the past year, developer usage has shifted sharply toward open-weight systems that teams can download, inspect, and run on their own infrastructure, and most of the leading options in that category now come from Chinese labs. The release is being covered accordingly, with one outlet describing it as the open-weight coding model pitched as the West's answer to DeepSeek and Qwen, noting that Poolside has released Laguna S 2.1, a 118-billion-parameter open-weight model built for agentic coding that the San Francisco startup says matches or exceeds models several times its size.

Poolside itself is candid that this isn't a frontier-topping release, but rather a meaningful step, saying Laguna S 2.1 is the first scale up from the XS model released a few weeks ago and while the model is not yet at the frontier, it's already very capable.

This is where the model gets genuinely interesting. Despite its modest active parameter count, Laguna S 2.1 goes toe-to-toe with — and in some cases beats — models many times larger. 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. On Terminal-Bench 2.1 and SWE-Bench Pro, Laguna S 2.1 matches or exceeds models several times its size, including DeepSeek-V4-Flash, NVIDIA's Nemotron 3 Ultra or Thinking Machines' Inkling.

But it's not the top of the leaderboard, and outside analysis is careful to note that Tencent Hy3 edges it on Terminal-Bench, and Poolside's own table shows Kimi K3 and Claude Fable 5 both scoring notably higher on the same tests. The fair summary from one analysis 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.

One of the most striking findings in Poolside's own release notes is just how much "thinking," or extended reasoning, boosts performance. Turning on max thinking lifts the model's score on Terminal-Bench 2.1 from 60.4% to 70.2% and on DeepSWE from 16.5% to 40.4% — a jump that's especially dramatic on the harder benchmark. Poolside also points out that some benchmarks are getting saturated at the top, meaning as the frontier advances, top scores cluster together and models that behave very differently end up no more than a few points apart. DeepSWE, by contrast, has more room to differentiate models, with frontier models ranging widely and some much larger open models scoring far lower.

Beyond the benchmark tables, Poolside highlighted a specific result meant to showcase long-horizon reasoning capability: they published a complete trajectory of Laguna S 2.1 independently proving a well-known unsolved combinatorics problem (Erdős Problem #397) that until recently only the largest frontier reasoning models had resolved.

Laguna S 2.1 is genuinely open. It ships under a license described by one analysis as permitting free commercial and non-commercial use and modification, a meaningfully more permissive posture than research-only or "community" licenses used by some other labs. You can get it through Hugging Face, through OpenRouter for free, through Poolside's own API, and through third-party coding tools that have already integrated it.

Laguna S 2.1 isn't claiming to be the best coding model in the world — a couple of rivals clearly outscore it on headline benchmarks. But that's arguably not the point. The story here is efficiency and access: a relatively small model holding its own against systems with far more compute behind them, released with a genuinely permissive license, small enough to self-host on modest hardware, and coming from a Western lab at a moment when most of the strongest open-weight options have been coming out of China. For enterprises, governments, and regulated industries that need to keep code and data in-house, that combination of permissive license, compact footprint, and competitive benchmarks is arguably a bigger deal than topping a leaderboard.