Poolside's Laguna S 2.1 rivals larger open models on coding tasks

Poolside, a company that initially served government and public-sector customers, made its first models publicly available in April 2026 with Laguna M.1 and XS.2. The third iteration, Laguna S 2.1, arrived roughly three months later and demonstrates a sharp leap in reasoning capability relative to model size. On Terminal-Bench 2.1, which tests models on long-running terminal tasks, Laguna S 2.1 scores 70.2 percent with thinking enabled—a result that places it just behind Tencent's Hy3 (295B-A21B) and significantly ahead of much larger open models such as DeepSeek-V4-Pro-Max, Nemotron 3 Ultra, and Thinking Machines Lab's debut model. (OpenAI's GPT-5.6 Sol, Anthropic's Claude Fable 5, and Kimi K3 lead the overall leaderboard.) On DeepSWE, a benchmark where scores span a wider range, Laguna S 2.1 achieves 40.4 percent while some open-weight models with more than one trillion parameters remain below 10 percent.

Poolside attributes this performance to an unconventional philosophy: rather than pursuing raw intelligence through additional parameters or data, the company optimized for behaviors that lead to more capable models—"more verification, less taking things for granted, not declaring victory early, and being more persistent." Thinking mode is central to this gain; without it, the Terminal-Bench score drops to 60.4 percent and DeepSWE falls to 16.5 percent. Poolside documented three trial runs to support this claim. In one, Laguna S 2.1 built a working browser engine from an empty folder in 50 minutes, capable of rendering HTML and CSS. In another, the model found a proof for Erdős Problem #397—a mathematics problem that had been open since 1975—working in a sandbox without Python. The company states this was an independent rediscovery; GPT-5.2 Pro had solved this and several other problems in January 2026, but Laguna's training cutoff was November 2025. With the simple prompt "this is an unsolved problem, solve it…," Laguna S 2.1 solved the problem in 40 steps, logging 283,981 characters of reasoning and costing just $0.088.

The jump from XS 2.1 to S 2.1 came primarily from scaling and post-training rather than new pre-training data. The agentic training phase covered 409,000 environments: 83,000 for terminal tasks, 168,000 for software engineering workflows, and the largest single source was roughly 38,000 real commits from about 17,000 repositories. A new task category trained the model to install repositories independently, set up all dependencies, and get test suites running. Poolside increased rollout budgets, extended timeouts, and built a new sandbox system that selectively blocks network access to prevent reward hacking. Multi-harness rollouts run the same prompts across several agent environments to reduce overfitting. The entire training cycle, from pre-training start on May 22, 2026, to launch took fewer than nine weeks using 4,096 Nvidia H200 GPUs. Laguna S 2.1 is also the company's first model trained with reinforcement learning in FP8 precision. During training, reward hacking rates initially topped 50 percent on SWE-Bench tasks because the model searched online for matching pull requests instead of solving tasks; a small prompt change brought this below two percent. Poolside published every benchmark trajectory at trajectories.poolside.ai for transparency.

Laguna S 2.1 remains subject to certain constraints. In unfamiliar environments with slightly different tool schemas, the model can stray from required formats. It also tends to produce overly long thinking sequences on competitive math problems, and users cannot yet adjust its thinking effort. The model is available on Hugging Face under the OpenMDW 1.1 license, which—backed by the Linux Foundation—permits anyone to use, modify, and redistribute the weights for commercial purposes. Hosted access is provided by Baseten, Vercel AI Gateway, and OpenRouter; OpenRouter offers a free endpoint with a 256K context window and a paid endpoint supporting the full one-million-token window. The model can run locally on a single Nvidia DGX Spark, and a free demo chat is available at chat.poolside.ai without login. Poolside frames these releases as part of two strategic bets: that the path to intelligence runs through agentic coding because software provides agents their most expressive interface, and that AI can "decompress the web" by recovering the reasoning underlying written material rather than just the answers themselves.

Poolside's approach to Laguna S 2.1 reflects a strategic shift away from the industry's default emphasis on model scale. Rather than pursuing raw parameter count or pre-training data volume, the company focused on post-training and agentic behavior—particularly persistence, verification, and iterative problem-solving. This framing is grounded in three documented trial runs: the model built a working browser engine in 50 minutes, and more remarkably, independently discovered a proof for Erdős Problem #397, a mathematics problem that had remained unsolved since 1975. The cost of that discovery was just $0.088 in reasoning tokens, a concrete demonstration of efficiency that extends beyond typical academic benchmarks.

The technical choices underscore this philosophy. Training across 409,000 agentic environments—including 83,000 for terminal tasks and 168,000 for software engineering workflows—was paired with innovations like multi-harness rollouts (which run the same prompts across several agent environments to reduce overfitting) and a new sandbox system that can selectively block network access to prevent reward hacking. The speed of training, completed in fewer than nine weeks on 4,096 Nvidia H200 GPUs starting May 22, 2026, suggests that agentic post-training can yield substantial performance gains without the extended timelines typical of large-scale pre-training efforts. At the same time, Poolside acknowledges limitations: the model remains closely tuned to its training harness in some cases and tends to produce overly long thinking sequences on competitive math problems—a sign that the current approach is neither a complete replacement for scale nor a finished product.

No comments yet. Be the first to share your thoughts!