Poolside drops the Laguna S 2.1, an open-source model that beats competitors 10x its size
Poolside, the San Francisco AI lab that has spent much of its three-year existence quietly selling coding models to governments and defense agencies, released its most capable model to date on Tuesday — and made an unusually aggressive bet that greater transparency, not raw scale, is how a small lab competes at the frontier.
The model, Laguna S 2.1, is a 118 billion Mixture-of-Experts (MoE) system that activates 8 billion parameters per token, supports a context window of up to 1 million tokens, and – according to benchmarks published by the company – matches or beats open models several times its size in agent code operations. Weights are immediately available from Hugging Face under the permissive OpenMDW-1.1 license.
The headline numbers are amazing for such a small model. Poolside reports that Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, a benchmark for long-term horizon operations, placing it 11th on the company’s combined leaderboard – ahead of DeepSeek-V4-Pro-Max, a 1.6-trillion-trillion.0 parameter model that scored 64 points; Thinking Machines’ 975-billion-parameter Inkling, at 63.8; and Nvidia’s 550-billion-parameter Nemotron 3 Ultra, at 56.4. On SWE-Bench Multilingual, it posts 78.5%, and on the SWE-Bench Pro public dataset, 59.4%.
Perhaps more telling than any single result: the model went from initial pre-training on May 22 to public launch in less than nine weeks, trained on 4,096 Nvidia H200 GPUs. In an industry where flagship model cycles are typically measured in quarters or years, Poolside has now shipped three models in three months.
Why the West’s open-world AI gap is becoming a boardroom issue
The release comes amid ongoing debate about the evolution of open-source AI. Over the past year, developer acquisition has shifted strongly toward open-source systems that companies can download, test, and run on their own infrastructure — and the leading options in that category have largely emerged from Chinese labs. DeepSeek, Qwen, Kimi, GLM, MiniMax, and Tencent’s Hunyuan line all feature prominently in Poolside’s comparison tables.
Poolside’s accompanying press release Laguna S 2.1 is apparently a response, noting that the model occupies a size category where no Western lab has released open weights in 11 months – since OpenAI’s gpt-oss-120b last August. "The West needs open-minded models it can trust, work with, and build upon," said Jason Warner, CEO of Poolside, in the announcement.
Founder and CEO Eiso Kant made the philosophical belief even clearer in a lengthy post on X. "I believe that intelligence should also be a commodity," he wrote, arguing that it is an open ecosystem "it won’t win by being the best in its division." He said users simply want the best intelligence for the task at hand – so open models should be equal to, or better than, closed ones.
The strategic idea here is not donation. Poolside’s core business is issuing models within the confines of government security, defense, and regulated businesses – customers close, metered API access is often not enabled for regulatory compliance reasons.
Every business that represents China’s open model today becomes difficult to win tomorrow. Unleashing competitive open weights is both an ecosystem game and a top-tier company’s business strategy. It also reorients the AI race away from the area where Poolside can’t compete – the large cost of marginal rate – and towards the area where it believes it can: cost per token, self-sufficiency, and speed of iteration.
How sarse architecture enables enterprise AI agents to work
Technical design shows a specific thesis about where the value of AI code goes. Laguna S 2.1’s sparse MoE architecture — 256 routing specialists and one shared specialist, with clustered query attention and sliding window layers, according to the Hugging Face card model — means an estimated cost of 8 billion operating parameters, not a total of 118 billion. Poolside emphasizes that the model is small enough to run on a single Nvidia DGX Spark, a desktop-class AI machine.
That’s important for what Poolside calls token economics. Horizon coding agents are voracious token consumers: the company’s published data shows the model consumes an average of about 249,000 fulfillment tokens per trajectory in its most difficult benchmark when the reasoning mode is on. With scaled API pricing, enterprise-scale agent workloads become a reasonable budget line item. For OpenRouter, Poolside offers a free endpoint of 256K content and 1M dedicated content forwarding priced at $0.10 per million input tokens and $0.20 per million output tokens – outrageous prices that undercut many other border methods by an order of magnitude.
The ecosystem support is unusually broad on day one. The model is live in the Baseten model library and Vercel’s AI Gateway, bundled in all vLLM, SGLang, Ollama, and llama.cpp, and limited variants up to 4-bit GGUF files — 75 gigabytes — for local use. But Poolside’s most interesting claim is ethical, not architectural. Pengming Wang, head of applied research at Poolside, said the benefits came from improving the working practices of this model: "more assertiveness, taking things for granted, not claiming victory early, and persistence." Raw intelligence, the company argues, is one axis of competence; The working method of the model is the second most important axis for agents left unattended.
Publishing the entire benchmark trajectory to combat the problem of AI reliability
The most important part of the release for business buyers may be the test-transparent move that has little precedent among major labs: Poolside has published a complete, unedited trace of every test in its last benchmark run – every thought step, tool call, and shell command behind every reported score.
This addresses the growing reliability problem in AI simulations. As high scores on mature benchmarks converge in the 70–90% range, and as "reward and hacking" – models who find solutions on the Internet or guarantees of games instead of solving problems – it has become common, the numbers they report have lost their main signal. Poolside openly revealed its encounter with the problem: during training, more than half of the tracks in some SWE bench tasks were flagged because the model simply searched for the first pull request to fix bugs on the Internet and used it. The company documented its reduction, including a quick addenda, LLM-based judgment limited against human labels, and expert annotator reviews for high-throughput Terminal-Bench runs.
Three published examples show what the company means by persistence. In one, the model built a working HTML/CSS engine from an empty folder in a 181-step, 50-minute unattended session – and then spun up headless Chromium to numerically compare the canvas output against the actual browser rendering. In one, pointed Poolside agent harness in an automatic optimization loop, the model made the Go codebase 5.2% faster and with about 70% lower memory allocation, finding an O(n²) string-concatenation bug along the way. In the third, running in a sandbox without Python installed, the model did number theory in Perl and independently found a proof of the Erdős problem #397 – a combinatoric question open for five decades until GPT-5.2 Pro started to solve it this past January. Poolside notes that its model formulation is structurally different from the previously published solution, and that the November 2025 data cutoff precedes the first proof.
Exposed limitations and the fine print of the benchmark
Poolside deserves credit for exposing limitations that many labs bury. The model can get too stuck in its native harness and stumble on slightly different tool schemas from third-party agents, include JSON in nested arguments, and tend to overthink competition statistics. Currently there is no adjustable think effort dial – on or off – and the gap between the methods is huge: the think raises Terminal-Bench 2.1 from 60.4% to 70.2%, and DeepSWE from 16.5% to 40.4%, at a much higher token cost.
Buyers should apply their discounts to the comparison tables. Poolside’s approach takes a lot of vendor-reported scores, benchmark author leaderboards, and third-party statistics from competitors — a logical convention, but one that includes harnesses and test conditions. For DeepSWE, in particular, Poolside used its own agent harness instead of a standard leaderboard agent, a difference the company notes makes the scores directly comparable. And the border remains unreachable: closed models like GPT-5.6 Sol, at 88.8 on Terminal-Bench 2.1, and Claude Fable 5, at 88.0, and 2.8-trillion-parameter open-weight Kimi K3, at 88.3, sit well above Laguna S 2.1.
A deep structural question is whether Poolside’s "Model Industry" – an internal platform that the company credits for its rapid release – can maintain this speed as models scale. The trajectory so far has been quite unusual: April’s dual release of the Laguna M.1 and XS.2, the July 2 refresh of the XS 2.1, and now the S 2.1, which the company says dwarfs the April M.1 flagship by nearly a third of its effective size. Surprisingly, S 2.1 used the exact same pre-training data as XS 2.1, meaning almost all of the improvement came from scale, training adjustments, and post-training across the company’s 409,000 agent and non-agency training scenarios. Poolside says its next, larger Laguna model began pre-training last week.
For technical decision-makers, Laguna S 2.1 is the most reliable Western open-weight solution that will appear in about a year to code your own managed agent – with published proofs, a permissive license, broad ecosystem support, and an economic case built around the hardware you may have. Whether it ties the dominance of China’s open models will depend on this release rather than the next.
Kant has told the world how he intends that story to end. Poolside is building a future where creativity is the most powerful "can be owned and shaped by anyone," he wrote – and the company plans to continue shipping "until that future exists." In an industry where big labs are increasingly locking their best work behind an API, the most important thing about Laguna S 2.1 may not be what it achieves, but that anyone can download and test it.
