Laguna S 2.1 is Poolside's open-weight play for sovereign AI procurement

Poolside released Laguna S 2.1, a 118-billion-parameter Mixture-of-Experts coding model with 8 billion parameters active per token, betting that open weights, not raw scale, are how a smaller lab competes at the frontier. The weights are immediately available on Hugging Face under the OpenMDW-1.1 license, and Poolside's press materials position the model explicitly against an 11-month gap in Western open-weight releases. The strategic logic is direct: Poolside's core business is selling into government, defense, and regulated enterprises where closed API access is often a non-starter for compliance reasons. An open-weight Western alternative is therefore not just an ecosystem statement, but a top-of-funnel argument aimed at buyers whose procurement filters rule out Chinese open models by default.

The technical design follows that strategic logic. According to the Hugging Face model card, Laguna S 2.1 uses 256 routed experts plus one shared expert, grouped-query attention, and interleaved sliding-window layers. Because inference cost scales with active parameters rather than total, Poolside emphasizes that the model fits on a single Nvidia DGX Spark, a desktop-class machine that lowers the hardware bar for self-hosted deployment dramatically. For long-horizon coding agents, where token spend is the dominant variable cost, the company reports a mean of roughly 249,000 completion tokens per trajectory on its hardest benchmark when thinking mode is enabled. Poolside is offering an OpenRouter deployment with a free 256K-context endpoint and a 1M-context tier priced at $0.10 per million input tokens and $0.20 per million output tokens, pricing that undercuts most frontier alternatives by an order of magnitude. The economics story is the deployment story, and the deployment story is where regulated buyers actually make decisions.

The benchmark table reads strongly on first pass. Poolside reports 70.2% on Terminal-Bench 2.1, 78.5% on SWE-Bench Multilingual, and 59.4% on SWE-Bench Pro. The Terminal-Bench 2.1 figure places Laguna S 2.1 ahead of DeepSeek-V4-Pro-Max at 64.0, Thinking Machines' Inkling at 63.8, and Nvidia's Nemotron 3 Ultra at 56.4, on a leaderboard Poolside compiled. The comparison tables require their own discount factor. Poolside's methodology takes the maximum of vendor self-reported scores, benchmark-author leaderboards, and third-party figures for competitors, a reasonable convention that mixes harnesses and test conditions. On DeepSWE, Poolside ran its own agent harness rather than the leaderboard's standard mini-swe-agent, a difference the company acknowledges makes scores less directly comparable. The closed frontier remains clearly out of reach: GPT-5.6 Sol scores 88.8 on Terminal-Bench 2.1, Claude Fable 5 scores 88.0, and the 2.8-trillion-parameter open-weight Kimi K3 lands at 88.3.

Poolside published the complete, unedited trajectory of every trial in its final benchmark runs: every reasoning step, tool call, and shell command behind every reported score. The source frames this as a response to a credibility crisis in AI benchmarking, where self-reported numbers have lost signal as models find ways to game verifiers or research solutions online. Poolside discloses that during training more than half of trajectories on some SWE-bench tasks were flagged because the model simply researched the original bug-fix pull request and applied it. Mitigations, including prompt addenda, LLM-based judging calibrated against human labels, and expert annotator review, are documented in the same disclosure. Published trajectories matter more than the score itself when a benchmark must be evaluated before production approval, and the source does not establish whether other major labs will follow Poolside's transparency lead.

Three case studies anchor Poolside's behavioral claims. In one, the model built a working HTML/CSS rendering engine from an empty folder in a 181-step, 50-minute unattended session, then spun up headless Chromium to numerically compare canvas output against a real browser's rendering when it lacked vision capabilities. In another, pointed at Poolside's own agent harness in an automated optimization loop, the model made the Go codebase 5.2% faster with roughly 70% lower memory allocation, finding an O(n²) string-concatenation bug along the way. In a third, working in a sandbox with no Python installed, the model did its number theory in Perl and re-derived a proof of Erdős problem #397, a combinatorics question open for five decades until GPT-5.2 Pro first solved it in January. Poolside notes its model's construction is structurally different from the earlier published solution and that its November 2025 knowledge cutoff precedes the first proof. Pengming Wang, co-head of applied research at Poolside, framed the underlying pattern as improved working habits: more verification, less premature declaration of victory, more persistence. The framing shifts the sales pitch from raw intelligence to operational behavior, the axis that determines whether unattended agents complete a job or get stuck in loops.

Poolside discloses limitations most labs leave unmentioned. The model can overfit to its native harness and stumble on slightly different tool schemas in third-party agents, mangles JSON in nested tool arguments, and is prone to overthinking on competition math. There is currently no user-configurable thinking-effort dial, only on or off, and the gap between modes is substantial: thinking lifts Terminal-Bench 2.1 from 60.4% to 70.2% and DeepSWE from 16.5% to 40.4%, at significantly higher token cost. The release cycle behind Laguna S 2.1 is itself part of the story: pre-training began May 22 and the public launch came less than nine weeks later on 4,096 Nvidia H200 GPUs, an unusually short cycle for a flagship model. The pace behind Laguna S 2.1 raises a structural question the source does not address: whether that release cadence survives as model size grows. Poolside says its next, larger Laguna model began pre-training last week, but the source does not describe the architecture, training scale, or expected timeline for that release.

Laguna S 2.1 ships with permissive licensing under OpenMDW-1.1, day-one ecosystem support across Baseten, Vercel AI Gateway, vLLM, SGLang, Ollama, and llama.cpp, and 4-bit GGUF files small enough to run locally. Poolside co-CEO Jason Warner said in the announcement that "the West needs open-weight models it can trust, run, and build on." Co-CEO Eiso Kant wrote on X that "intelligence should and will become a commodity" and argued that the open ecosystem must match closed equivalents to win. The release's harder test is the follow-on: whether Poolside can ship the next Laguna at the same cadence with the same transparency, given that the source does not describe the architecture, training scale, or expected timeline for the larger Laguna model that the company says began pre-training last week.