The 118-billion-parameter model matches models several times its size on agentic coding and is small enough to run on a single desktop. It is available today under an OpenMDW-1.1 license.
SAN FRANCISCO, July 21, 2026 (GLOBE NEWSWIRE) -- Poolside today released Laguna S 2.1, a 118-billion-parameter open-weight foundation model built for agentic coding. On Terminal-Bench 2.1 and SWE-Bench Pro, Laguna S 2.1 matches or exceeds models several times its size, including DeepSeek-V4-Flash, NVIDIA's Nemotron 3 Ultra or Thinking Machines' Inkling. It is small enough to run on a single NVIDIA DGX Spark, and the weights are available today on Hugging Face under an OpenMDW-1.1 license. Laguna S 2.1 is the first scale up from the XS model released a few weeks ago and while the model is not yet at the frontier, it's already very capable.
Laguna S 2.1 arrives as the question of who supplies the West’s open-weight models has moved from research circles into boardrooms and Washington. Over the past year, developer usage has shifted sharply toward open-weight systems that teams can download, inspect, and run on their own infrastructure, and most of the leading options in that category now come from Chinese labs. As several US labs have moved their most capable systems behind closed APIs and government-gated rollouts, the West has been left without a competitive open-weight option. Laguna S 2.1 is built to be that option, and to compete with the incumbents on the terms buyers now care about: cost, control, and the ability to keep data in house.
The release also reframes a debate that has centered on frontier capability and the capital required to reach it. As enterprises move from unconstrained token spending toward a focus on efficiency and return, the operative questions have become the price of running intelligence at scale, whether that intelligence can be self-hosted for sovereignty and compliance, and whether the business behind a model is sustainable. Laguna S 2.1 is designed for exactly those constraints.
THE EVIDENCE
Poolside is releasing Laguna S 2.1 alongside a key set of agentic coding evaluations run using pool, Poolside’s agent harness, so researchers and developers can reproduce the results on their own hardware.
| Model | Size | Terminal- Bench 2.1 | SWE-bench Multilingual | SWE-Bench Pro (Public Dataset) | DeepSWE v1.1 |
| Laguna S 2.1 | 118B-A8B | 70.2% | 78.5%* | 59.4% | 40.4% |
| Laguna M.1 (April 2026) | 225B-A23B | 44.9% | 63.1% | 49.2% | - |
| Tencent Hy3 | 295B-A21B | 71.7% | 75.8% | 57.9% | - |
| Thinking Machines Inkling | 975B-A41B | 63.8% | - | 54.3% | - |
| Nemotron 3 Ultra | 550B-A55B | 56.4% | 67.7% | - | - |
| MiniMax M3 | 428B-A23B | 66.0% | - | 59.0% | - |
| DeepSeek V4-Flash Max | 284B-A13B | 61.8%** | 73.3% | 52.6% | - |
| DeepSeek V4-Pro Max | 1.6T-A49B | 64.0%** | 76.2% | 55.4% | 9.0% |
| Qwen3.7-Max | - | 74.5%** | 78.3% | 60.6% | - |
| Claude Sonnet 5 | - | 80.4% | - | 63.2% | 54.0% |
| GPT-5.6 Luna Max | - | 82.5% | - | - | 67.0% |
Comparison scores use the highest publicly referenced results for each model. Full methodology, sampling parameters, and per-benchmark detail will accompany the release.
** As reported by Artificial Analysis.
As a further demonstration of long-horizon reasoning, Poolside publishes a complete trajectory of Laguna S 2.1 independently proving Erdős Problem #397, a combinatorics problem that until recently only the largest frontier reasoning models had resolved.
TOKEN ECONOMICS
Running Laguna S 2.1 locally changes the arithmetic of deploying coding agents. Small enough to run on a single DGX Spark, it lets teams move high-volume agentic work off metered APIs and onto hardware they control. For the government, defense, and other highly regulated organizations that Poolside already serves, self-hosting keeps sensitive code and data inside their own environment. Laguna S 2.1 also occupies a size class that no Western lab has released an open-weight model into in 11 months, since gpt-oss-120b’ release in August last year.
HOW WE BUILD
Laguna S 2.1 is the latest model out of Poolside’s Model Factory, the company’s internal platform for training, evaluating, and iterating on foundation models. The Factory automates the work that traditionally makes model development slow, from architecture ablations to reinforcement learning from code execution, running experiments that once took weeks in under an hour. It is what lets Poolside ship on a roughly five-week cadence. For reference, Laguna S 2.1 took less than 4 weeks to train end to end on 4,000 H200 GPUs.
“The West needs open-weight models it can trust, run, and build on,” said Jason Warner, co-CEO of Poolside. “Laguna S 2.1 is our answer. It is a model that enterprises and governments can put into production today, on their own hardware, at a cost that makes agentic coding practical at scale.”
“Laguna S 2.1 does the work of models several times its size because of how we build, not despite it,” said Eiso Kant, co-founder and co-CEO of Poolside. “The Model Factory automates the work that traditionally makes model development slow and error prone, allowing its researchers to run an order of magnitude more experiments.”
AVAILABILITY
Laguna S 2.1 is available today. Weights are on Hugging Face under an OpenMDW-1.1 license. The model is also available through Poolside’s API and on OpenRouter, and runs in pool, Poolside’s agent harness.
→ Download weights on Hugging Face
ABOUT POOLSIDE
Poolside is a foundation AI lab building open-weight models for software engineering on the path to human-level artificial intelligence. Founded in 2023 and headquartered in San Francisco, Poolside trains its models from scratch on its own data and infrastructure, using reinforcement learning from code execution feedback so its models improve from their own successes and failures, the way human developers do. Poolside deploys directly into the development environments of government, defense, and other highly regulated organizations. Learn more at https://poolside.ai.
Media contact: press@poolside.ai
