PoolsideA San Francisco AI lab that has spent most 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 how a smaller lab competes on the frontier is radical transparency, not raw scale.
model, Laguna S 2.1118 billion parameters Mix of Experts (MC) system enables a total of 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 coding tasks. It’s weights available immediately At Hugging Face under the permissive OpenMDW-1.1 license.
Headline numbers are surprising for such a small model. “Poolside” reported on this Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1is the benchmark for long-horizon terminal tasks, placing it 11th on the company’s compiled leaderboard — ahead. DeepSeek-V4-Pro-Max1.6 trillion parameter model scoring 64.0; 975 billion parameters of Thinking Machines Complex63.8; and Nvidia’s 550 billion settings Nemotron 3 Ultra56.4. Active SWE-Bench Multilingual78.5% writes and continues SWE-Bench Propublic databases, 59.4%.
Perhaps more telling than any score: the model went from initial training on May 22 to public release 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 have the West’s open-weight AI gaps become a boardroom challenge?
Its release is in the midst of an increasingly heated debate origin of AI in open weight. Over the past year, developer adoption has shifted decisively toward open-weight systems that companies can download, test, and run on their own infrastructure, and the leading choices in this category have come overwhelmingly from Chinese labs. DeepSeek, Gwen, Like, GLM, MiniMaxand Hunyuan from Tencent List all features prominently in Poolside’s own comparison charts.
Poolside press release frames Laguna S 2.1 in clear response, he pointed out that the model occupies a size class that no Western lab has released open weights for 11 months – since OpenAI gpt-oss-120b last August. "The West needs open models it can trust, operate and build upon." said Jason Warner, CEO of Poolside.
Co-founder and CEO Eiso Kant further clarified the philosophical stakes. long post On X. "I believe that intelligence should and will be commoditized." wrote arguing that there is an open ecosystem "it won’t win by being the best in its category." Users, he argued, simply want the best intelligence for the task at hand—so open models must be equal to or better than their closed equivalents.
The strategic logic here is not charity. Poolside’s core business is to deploy models within the security boundaries of government, defense and regulated enterprises – for whom closed, metered API access is often a non-starter for compliance and sovereignty reasons.
Any enterprise that standardizes on China’s open model today will find it difficult to win tomorrow. Launching competitive open weights is a top-tier strategy for both the ecosystem play and the company’s highly secure deployment business. It also steers the AI race away from areas where Poolside can’t compete—borderline capital expenditures—and toward where it believes it can: cost per token, self-hosting, and iteration speed.
How a sparse architecture makes it convenient to manage enterprise AI agents
Technical design reflects a specific thesis about where value moves in coding artificial intelligence. Laguna S 2.1’s sparse TN architecture — 256 directed experts plus one common expert, with clustered query focus and interleaved sliding window layers Hugging Face model card — means the scale of inference costs with 8 billion active parameters, not just 118 billion. Poolside points out that the model is small enough to run an Nvidia DGX Spark, a desktop-class AI machine.
This is important for what Poolside calls the token economy. Over the long horizon, coding agents are voracious consumers of tokens: the company’s published data shows that the model consumes about 249,000 completion tokens per trajectory in the most difficult benchmark with think mode enabled. At scaled API prices, enterprise-wide agent workloads become a meaningful budget item. On OpenRouter, Poolside offers a free 256K context endpoint and a dedicated 1M context placement at $0.10 per million input tokens and $0.20 per million output tokens – an aggressive price that undercuts most edge alternatives by an order of magnitude.
Ecosystem support is unusually broad for a first day. The model is live Basenet’s model library and Verselin’s AI Gatewaywith integration between vLLM, SGLang, to beand call.cppplus quantized options up to 4-bit GGUF files for local use – 75 gigabytes. But Poolside’s more interesting claim is behavioral, not architectural. Pengming Wang, co-president of applied research at Poolside, said the gains came from improving the model’s work habits: "more checking, not taking things for granted, not declaring victory early and being more stubborn." The company claims that raw intelligence is an axis of capability; the way the model works is the second axis of great importance for agents who are left unattended for hours.
Publishing each benchmark trajectory to counter AI’s credibility crisis
For enterprise buyers, the most important part of the release may be a move of evaluation-transparency that has little precedent among major labs: Poolside has published the full, unedited trajectory of every test in recent benchmark tests – every reasoning step, tool call, and shell command behind every reported score.
This solves the problem of increased reliability in AI benchmarking. The highest scores on mature criteria are in the 70-90% range and so on "premium hacking" — models or game checkers who find solutions online instead of solving problems — have become endemic, with self-reported numbers losing much of their signal. Poolside publicly disclosed its own encounter with the problem: during training, more than half of the trajectories on some SWE bench tasks were flagged because the model simply examined and applied an online error correction query. The company has documented its effects, including quick additions, LLM-based judging calibrated against human tags, and expert annotator review of a high-scoring Terminal-Bench run.
Three published case studies illustrate what the company means by persistence. In one, the model built an HTML/CSS rendering engine running from an empty directory in a 181-step, 50-minute unsupervised session — then, because it had no vision, fired up headless Chromium to numerically compare its canvas output to a real browser’s rendering. In another, a model pointing to Poolside’s own agent harness in an automated optimization loop made the Go codebase 5.2% faster with about 70% less memory allocation, finding one O(n²) string-merging error along the way. Third, running in a sandbox without Python installed, the model did its number theory in Perl and independently re-obtained proof of Erdős problem #397—a combinatorial question that had been open for five decades until GPT-5.2 Pro solved it for the first time last January. Poolside notes that the construction of his model is structurally different from the previously published solution, and that his November 2025 data cutoff is before the first proof.
What the revealed limitations and comparative fine print reveal
Poolside deserves credit for revealing limitations that most labs have buried. The model can adapt to its native harnesses and stumble on slightly different tool schemas in third-party agents, handles JSON in internal tool arguments, and tends to think a lot about competition math. There is currently no user-configurable think-effort dial – only on or off – and the gap between modes is huge: increasing think-ability Terminal-Bench 2.1 from 60.4% to 70.2%, and DeepSWE From 16.5% to 40.4%, with a significantly higher token value.
Buyers must apply their own discount to comparison charts. Poolside’s methodology takes vendor-reported scores, benchmark-authored leaderboards, and third-party figures for competitors—a reasonable convention, but one that confounds trailers and test conditions. Active DeepSWEnotably, Poolside ran its own agent trailer rather than the leaderboard’s standard mini-swe-agent, a difference the company admits makes the scores directly comparable. The boundary remains clearly unattainable: as closed models GPT-5.6 Sol88.8 in Terminal-Bench 2.1 and Representation of Claude 5At 88.0, with an open weight of 2.8 trillion parameters Like K3At 88.3, sit well above the Laguna S 2.1.
A deeper structural question was Poolside’s "Model factory" — an in-house platform that the company credits for its rapid release cadence — can keep up that pace as models scale. The trajectory so far is truly extraordinary: the April dual launch of the Laguna M.1 and XS.2, the July 2 refresh of the XS 2.1, and now the S 2.1, which dwarfs the company’s April flagship M.1 by nearly a third of its asset size. Interestingly, S 2.1 used the same pre-training data as XS 2.1, meaning almost all of the improvements came from scaling, training corrections and post-training training across the company’s 409,000 agent and non-agent training environments. Poolside says the next, larger Laguna model began preparations last week.
For technical decision makers Laguna S 2.1 It’s the most reliable Western open-weight option to emerge in nearly a year for self-hosted agent coding, with published evidence, a permissive license, broad ecosystem support, and an economics story built around hardware you can own. Whether or not Chinese outdoor models dominate will depend less on this release than those who follow it.
Kant, for his part, has already told the world how he intends to end this story. Poolside is building towards a future of the most skilled minds "can possess and shape anyone," he wrote — and the company plans to resume shipping "until that future exists." In an industry where the biggest labs are increasingly locking their best work behind APIs, the most radical thing about Laguna S 2.1 may not be what it achieves, but anyone can download and check it out.
