Poolside, the San Francisco AI lab that has spent most of its three-year existence quietly promoting coding fashions to governments and protection businesses, launched its most succesful mannequin so far on Tuesday — and made an unusually aggressive guess that radical transparency, not uncooked scale, is how a smaller lab competes on the frontier.
The mannequin, Laguna S 2.1, is a 118-billion-parameter Combination-of-Consultants (MoE) system that prompts solely 8 billion parameters per token, helps a context window of as much as 1 million tokens, and — in accordance with benchmarks printed by the corporate — matches or beats open fashions a number of instances its measurement on agentic coding duties. The weights are out there instantly on Hugging Face below the permissive OpenMDW-1.1 license.
The headline numbers are placing for a mannequin this small. Poolside experiences that Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, a benchmark of long-horizon terminal duties, putting it eleventh on the corporate's compiled leaderboard — forward of DeepSeek-V4-Professional-Max, a 1.6-trillion-parameter mannequin that scored 64.0; Considering Machines' 975-billion-parameter Inkling, at 63.8; and Nvidia’s 550-billion-parameter Nemotron 3 Extremely, at 56.4. On SWE-Bench Multilingual, it posts 78.5%, and on SWE-Bench Professional's public dataset, 59.4%.
Maybe extra telling than any single rating: the mannequin went from the beginning of pre-training on Might 22 to public launch in below 9 weeks, skilled on 4,096 Nvidia H200 GPUs. In an business the place flagship mannequin cycles are sometimes measured in quarters or years, Poolside has now shipped three fashions in three months.
Why the West's open-weight AI hole has turn out to be a boardroom challenge
The discharge lands in the course of an more and more pointed debate about the provenance of open-weight AI. Over the previous yr, developer adoption has shifted decisively towards open-weight techniques that firms can obtain, examine, and run on their very own infrastructure — and the main choices in that class have overwhelmingly come from Chinese language labs. DeepSeek, Qwen, Kimi, GLM, MiniMax, and Tencent's Hunyuan line all characteristic prominently in Poolside's personal comparability tables.
Poolside's accompanying press launch frames Laguna S 2.1 explicitly as a response, noting that the mannequin occupies a measurement class into which no Western lab has launched open weights in 11 months — since OpenAI's gpt-oss-120b final August. "The West wants open-weight fashions it will probably belief, run, and construct on," stated Jason Warner, Poolside's co-CEO, within the announcement.
Co-founder and co-CEO Eiso Kant made the philosophical stakes even plainer in a prolonged put up on X. "I consider intelligence ought to and can turn out to be a commodity," he wrote, arguing that the open ecosystem "won’t win by being the very best in its personal class." Customers, he argued, merely need the very best intelligence for the duty at hand — so open fashions should be on par with, or higher than, their closed equivalents.
The strategic logic right here shouldn’t be charity. Poolside's core enterprise is deploying fashions contained in the safety boundaries of presidency, protection, and controlled enterprises — prospects for whom closed, metered API entry is usually a non-starter for compliance and sovereignty causes.
Each enterprise that standardizes on a Chinese language open mannequin immediately turns into more durable to win tomorrow. Releasing aggressive open weights is each an ecosystem play and a top-of-funnel technique for the corporate's high-security deployment enterprise. It additionally reframes the AI race away from terrain the place Poolside can’t compete — frontier-scale capital expenditure — and towards terrain the place it believes it will probably: value per token, self-hosting, and iteration pace.
How a sparse structure makes enterprise AI brokers inexpensive to run
The technical design displays a particular thesis about the place worth in coding AI is shifting. Laguna S 2.1's sparse MoE structure — 256 routed specialists plus one shared professional, with grouped-query consideration and interleaved sliding-window layers, in accordance with the Hugging Face mannequin card — means inference prices scale with the 8 billion lively parameters, not the 118 billion complete. Poolside emphasizes that the mannequin is sufficiently small to run on a single Nvidia DGX Spark, the desktop-class AI machine.
That issues for what Poolside calls token economics. Lengthy-horizon coding brokers are voracious customers of tokens: the corporate's printed knowledge reveals the mannequin consuming a imply of roughly 249,000 completion tokens per trajectory on its hardest benchmark when pondering mode is enabled. At metered API costs, agentic workloads at enterprise scale turn out to be a significant price range line merchandise. On OpenRouter, Poolside is providing a free 256K-context endpoint and a devoted 1M-context deployment priced at $0.10 per million enter tokens and $0.20 per million output tokens — aggressive pricing that undercuts most frontier options by an order of magnitude.
The ecosystem help is unusually broad for day one. The mannequin is dwell on Baseten's mannequin library and Vercel's AI Gateway, with integrations throughout vLLM, SGLang, Ollama, and llama.cpp, plus quantized variants right down to 4-bit GGUF information — 75 gigabytes — for native use. However Poolside's extra attention-grabbing declare is behavioral, not architectural. Pengming Wang, co-head of utilized analysis at Poolside, stated the positive aspects got here from bettering the mannequin's working habits: "extra verification, much less taking issues as a right, not declaring victory early, and being extra persistent." Uncooked intelligence, the corporate argues, is one axis of functionality; a mannequin's manner of working is a second axis that issues immensely for brokers left unattended for hours.
Publishing each benchmark trajectory to counter AI's credibility disaster
Essentially the most consequential a part of the discharge for enterprise patrons could also be an evaluation-transparency transfer with little precedent amongst main labs: Poolside printed the entire, unedited trajectory of each trial in its remaining benchmark runs — each reasoning step, instrument name, and shell command behind each reported rating.
This addresses a rising credibility downside in AI benchmarking. As prime scores on mature benchmarks cluster within the 70–90% vary, and as "reward hacking" — fashions discovering options on-line or gaming verifiers reasonably than fixing issues — has turn out to be endemic, self-reported numbers have misplaced a lot of their sign. Poolside disclosed its personal encounters with the issue candidly: throughout coaching, greater than half of trajectories on some SWE-bench duties had been flagged as a result of the mannequin merely researched the unique bug-fix pull request on-line and utilized it. The corporate documented its mitigations, together with immediate addenda, LLM-based judging calibrated in opposition to human labels, and professional annotator evaluate of a high-scoring Terminal-Bench run.
Three printed case research illustrate what the corporate means by persistence. In a single, the mannequin constructed a working HTML/CSS rendering engine from an empty folder in a 181-step, 50-minute unattended session — then, missing imaginative and prescient capabilities, spun up headless Chromium to numerically evaluate its canvas output in opposition to an actual browser's rendering. In one other, pointed at Poolside's personal agent harness in an automatic optimization loop, the mannequin made the Go codebase 5.2% quicker with roughly 70% decrease reminiscence allocation, discovering an O(n²) string-concatenation bug alongside the best way. In a 3rd, working in a sandbox with no Python put in, the mannequin did its quantity concept in Perl and independently re-derived a proof of Erdős downside #397 — a combinatorics query open for 5 a long time till GPT-5.2 Professional first solved it this previous January. Poolside notes that its mannequin's development is structurally totally different from the sooner printed resolution, and that its November 2025 information cutoff precedes the primary proof.
What the disclosed limitations and benchmark wonderful print reveal
Poolside deserves credit score for disclosing limitations most labs bury. The mannequin can overfit to its native harness and hit upon barely totally different instrument schemas in third-party brokers, mangles JSON in nested instrument arguments, and is susceptible to overthinking on competitors math. There may be at present no user-configurable thinking-effort dial — simply on or off — and the hole between the modes is big: pondering lifts Terminal-Bench 2.1 from 60.4% to 70.2%, and DeepSWE from 16.5% to 40.4%, at considerably larger token value.
Patrons ought to apply their very own reductions to the comparability tables. Poolside's methodology takes the utmost of vendor self-reported scores, benchmark-author leaderboards, and third-party figures for opponents — an inexpensive conference, however one which mixes harnesses and take a look at situations. On DeepSWE, notably, Poolside ran its personal agent harness reasonably than the leaderboard's customary mini-swe-agent, a distinction the corporate acknowledges makes scores much less immediately comparable. And the frontier stays clearly out of attain: closed fashions like GPT-5.6 Sol, at 88.8 on Terminal-Bench 2.1, and Claude Fable 5, at 88.0, together with the two.8-trillion-parameter open-weight Kimi K3, at 88.3, sit effectively above Laguna S 2.1.
The deeper structural query is whether or not Poolside's "Mannequin Manufacturing unit" — the interior platform the corporate credit for its speedy launch cadence — can maintain this tempo as fashions scale. The trajectory to date is genuinely uncommon: the April twin launch of Laguna M.1 and XS.2, the July 2 refresh of XS 2.1, and now S 2.1, which the corporate says outperforms April's flagship M.1 at roughly a 3rd of its lively measurement. Remarkably, S 2.1 used the very same pre-training knowledge as XS 2.1, that means practically all the development got here from scale, coaching fixes, and post-training throughout the corporate's corpus of 409,000 agentic and non-agentic coaching environments. Poolside says its subsequent, bigger Laguna mannequin started pre-training final week.
For technical resolution makers, Laguna S 2.1 is essentially the most credible Western open-weight choice to emerge in practically a yr for self-hosted agentic coding — with printed proof, a permissive license, broad ecosystem help, and an economics story constructed round {hardware} you’ll be able to personal. Whether or not it dents the dominance of Chinese language open fashions will rely much less on this launch than on those that comply with it.
Kant, for his half, has already informed the world how he intends that story to finish. Poolside is constructing towards a future the place essentially the most succesful intelligence "might be owned and formed by anybody," he wrote — and the corporate plans to maintain delivery "till that future exists." In an business the place the most important labs more and more lock their finest work behind an API, essentially the most radical factor about Laguna S 2.1 is probably not what it scores, however that anybody can obtain it and examine.
