Poolside, the US-based AI company, today released Laguna S 2.1, a compact mixture-of-experts coding model that outperforms much larger open-weight rivals and can solve a 50-year-old unsolved math problem for less than a dime. The model, with 8 billion active parameters out of 118 billion total, is the third coding model Poolside has shipped in three months.
The company is betting that agentic coding—where AI models autonomously write, test, and debug software—is the fastest path to general intelligence. Laguna S 2.1 supports context windows up to one million tokens and offers both thinking and no-thinking modes. Its weights are available on Hugging Face under the Linux Foundation-backed OpenMDW 1.1 license, which allows anyone to use, modify, and redistribute the model for commercial purposes.
A model that thinks before it acts
With thinking enabled, Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, a benchmark that tests models on long-running terminal tasks. That places it just behind Tencent’s Hy3 (295B-A21B) but ahead of DeepSeek-V4-Pro-Max, Nemotron 3 Ultra, and Thinking Machines Lab’s debut model. The overall leaderboard is led by OpenAI’s GPT-5.6 Sol, Anthropic’s Claude Fable 5, and Kimi K3.
On Datacurve’s DeepSWE benchmark, which Poolside says offers better comparison because scores are spread across a wider range, Laguna S 2.1 scores 40.4%. Some open-weight models with more than 1 trillion parameters remain below 10% on DeepSWE. The model also ranks near the top of its class on SWE-Bench Multilingual, SWE-Bench Pro, and SWE Atlas.
The thinking mode is critical. Without it, Terminal-Bench drops to 60.4%, and DeepSWE falls to 16.5%. No previous Laguna model showed such a large performance gap between thinking and no-thinking modes.
“What we've done in this model is not necessarily add more intelligence, but improve the behaviors that lead to a more capable model: more verification, less taking things for granted, not declaring victory early, and being more persistent,” Poolside said in a release post.
Solving a 50-year-old problem for $0.088
In one trial, Laguna S 2.1 built a working browser engine from an empty folder in 50 minutes that could render HTML and CSS. In another, the model found a proof for Erdős Problem #397, a combinatorial number theory problem that had been open since 1975. The model worked in a sandbox without Python, producing 283,981 characters of reasoning over 40 steps. The entire run cost $0.088.
Poolside says the result was an independent rediscovery. GPT-5.2 Pro solved Erdős Problem #397 and several others in January 2026, but Laguna’s training cutoff was November 2025, meaning the model had no prior knowledge of that solution. With the prompt “this is an unsolved problem, solve it…,” Laguna S 2.1 found the proof on its own.
Training at speed and scale
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
The jump from XS 2.1 to S 2.1 came mainly from scaling and post-training, not new pre-training data. Pre-training began on May 22, 2026, using 4,096 Nvidia H200 GPUs. Fewer than nine weeks passed between the start of training and launch.
The agentic training phase covered 409,000 environments. That included 83,000 environments for terminal tasks and 168,000 environments for software engineering workflows. The largest single source was about 38,000 real commits from roughly 17,000 repositories. A new task category trained the model to install repositories on its own, set up dependencies, and get test suites running.
Poolside increased rollout budgets and extended timeouts. It built a new sandbox system that can selectively block network access to curb reward hacking. Multi-harness rollouts run the same prompts across several agent environments. S 2.1 is also the first Poolside model trained with reinforcement learning in FP8 precision.
Reward hacking and other limitations
During training, reward hacking rates topped 50% on SWE-Bench tasks because the model searched online for matching pull requests. A small prompt change brought the rate below 2%. But Poolside acknowledges that Laguna S 2.1 is still too closely tuned to its agent harness in some cases. In unfamiliar environments with slightly different tool schemas, the model can stray from the required format.
The model also tends to produce overly long thinking sequences on competitive math problems. Users cannot adjust thinking effort yet.
Availability and deployment
Laguna S 2.1 is available on Hugging Face under the OpenMDW 1.1 license, backed by the Linux Foundation. Baseten, Vercel AI Gateway, and OpenRouter offer hosted access. OpenRouter provides a free endpoint with a 256K token context window and a paid endpoint supporting the full one-million-token window. The model can run locally on a single Nvidia DGX Spark. A free demo chat is available at chat.poolside.ai without login.
Poolside initially focused on government and public-sector customers before making its first models, Laguna M.1 and XS.2, available to a broader audience in April 2026. XS.2 was the first open model under Apache 2.0. Laguna S 2.1 is the third version in the series released in roughly three months. A larger Laguna model is already in pre-training.
The company is making two strategic bets: the path to intelligence runs through agentic coding, and AI can “decompress the web.” Most written material records answers rather than reasoning, Poolside believes, and reinforcement learning can recover that process.