Poolside Releases Laguna S 2.1, an 8B-Active-Parameter Open Coding Model That Rivals Systems 20x Its Size

Poolside has released Laguna S 2.1, an open-weight coding model that the company says outperforms open-weight systems many times its size on agentic coding benchmarks. It is Poolside's third coding model release in roughly three months.

Laguna S 2.1 is a mixture-of-experts model with 118 billion total parameters and 8 billion active parameters per token. It supports context windows up to 1 million tokens and offers both thinking and no-thinking inference modes.

Benchmark performance

With thinking enabled, Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, a benchmark for long-running terminal tasks. According to Poolside, that places it just behind Tencent's Hy3 (295B-A21B) and ahead of larger open models including DeepSeek-V4-Pro-Max, Nemotron 3 Ultra, and Thinking Machines Lab's debut model. The top of that leaderboard is held by closed models: OpenAI's GPT-5.6 Sol, Anthropic's Claude Fable 5, and Kimi K3.

Poolside says Datacurve's DeepSWE benchmark is a more informative comparison because scores spread across a wider range. Laguna S 2.1 scores 40.4% on DeepSWE, while some open-weight models exceeding 1 trillion parameters score below 10%, according to the company. Poolside also claims strong placement on SWE-Bench Multilingual, SWE-Bench Pro, and SWE Atlas.

Thinking mode matters significantly for this model. Without it, Terminal-Bench score drops to 60.4% and DeepSWE falls to 16.5%. Poolside says this is the largest gap between thinking and non-thinking modes it has observed in any Laguna model to date. Users currently cannot adjust the thinking effort level.

Persistence over scale

Poolside frames the release around behavioral improvements rather than raw capability gains. "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," the company wrote in its release post. Earlier Laguna models reportedly stopped after partially passing test suites or abandoned working approaches prematurely.

Poolside says the jump from XS 2.1 to S 2.1 came primarily from post-training, not new pre-training data. The agentic training phase spanned 409,000 environments, including 83,000 terminal tasks and 168,000 software engineering workflows, with roughly 38,000 real commits from about 17,000 repositories as the largest single data source. A new sandbox system selectively blocks network access to curb reward hacking — during training, reward hacking rates on SWE-Bench tasks reportedly exceeded 50% before a prompt change brought that below 2%.

Pre-training started May 22, 2026, using 4,096 Nvidia H200 GPUs, with fewer than nine weeks between training start and launch, according to Poolside. This is the company's first model trained with reinforcement learning in FP8 precision.

As a demonstration, Poolside says Laguna S 2.1 built a working browser engine capable of rendering HTML and CSS from an empty folder in 50 minutes, and found a proof for Erdős Problem #397 — open since 1975 — in a sandbox without Python access, in 40 steps costing $0.088. OpenAI's GPT-5.2 Pro reportedly solved the same problem in January 2026; Laguna's training cutoff was November 2025, which Poolside describes as an independent rediscovery.

Availability

Laguna S 2.1 is available on Hugging Face under the OpenMDW 1.1 license, a Linux Foundation-backed license permitting commercial use, modification, and redistribution. Hosted access is offered through Baseten, Vercel AI Gateway, and OpenRouter, which provides a free endpoint at 256K context and a paid endpoint for the full 1-million-token window. Poolside says the model runs locally on a single Nvidia DGX Spark. A login-free demo is available at chat.poolside.ai. Pricing for hosted endpoints was not disclosed in Poolside's release.

Poolside acknowledges limitations: the model remains tightly coupled to its own agent harness and can deviate from required formats in unfamiliar tool schemas, and it tends to over-generate reasoning tokens on competitive math problems.

What this means

Laguna S 2.1 is a data point for an argument the industry has been testing all year: that post-training behavior — verification, persistence, refusal to declare premature victory — can close a meaningful gap with models many times larger, at least on agentic coding tasks. Poolside's decision to publish full benchmark trajectories at trajectories.poolside.ai gives outside researchers a way to check these claims rather than take them on faith, which is more transparency than most labs offer.

The more consequential number here may be the reward-hacking rate during training — over 50% on SWE-Bench before mitigation. That suggests current RL training pipelines for coding agents are still fragile to exploitation, and that benchmark scores across the industry deserve scrutiny for similar failure modes. A larger Laguna model is already in pre-training, according to Poolside, which will be the next real test of whether persistence-focused post-training scales alongside parameter count or is a one-time efficiency gain.