Poolside has released Laguna S 2.1, a 118-billion-parameter open-weight model built for agentic coding that the San Francisco startup says matches or exceeds models several times its size. The model uses a mixture-of-experts architecture with eight billion active parameters per token, is compact enough to run on a single Nvidia DGX Spark desktop system, and the weights are available on Hugging Face under the Linux Foundation's OpenMDW license.

On Terminal-Bench and SWE-Bench Pro, two agentic coding evaluations, Laguna S 2.1 scored just over 70 percent and nearly 60 percent respectively, matching or beating models from DeepSeek, Nvidia, and Thinking Machines that carry two to eight times as many active parameters. Poolside acknowledges the model is “not yet at the frontier,” with closed-source systems from OpenAI and Anthropic still scoring well above it on the same benchmarks.

The release is framed as a direct response to the dominance of Chinese labs in the open-weight category, where DeepSeek, Alibaba's Qwen family, and Moonshot's Kimi have set the pace for more than a year. No Western lab had released an open-weight model in the 118-billion-parameter class for 11 months before this launch, according to the company. Forbes reported that Poolside explicitly positioned the release as an effort to give Western enterprises and governments a self-hosted alternative they can run without sending data to a foreign provider.

Poolside was founded in 2023 by Jason Warner, formerly chief technology officer at GitHub, and Eiso Kant, and raised $500 million in a Series B in October 2024 at a $3 billion valuation with backing from Nvidia and eBay. A planned $2 billion Series C that would have valued the company at $14 billion collapsed in April 2026 after CoreWeave walked away from a joint data centre project in Texas. The company now serves government, defense, and other highly regulated organizations through its API and agent harness.

Poolside says it built the model using its internal Model Factory platform, which automates architecture search and reinforcement learning from code execution, and completed training in under four weeks on 4,000 Nvidia H200 GPUs. The smaller Laguna XS launched three weeks earlier, and the company says it ships new models on roughly a five-week cadence. As a demonstration of long-horizon reasoning, Poolside published a trajectory of the model independently solving a combinatorics problem that until recently only the largest frontier models had resolved.

The bet is that enterprises will pay to run a capable coding model on their own hardware rather than send prompts to a closed API, a thesis that depends on Laguna S 2.1 performing in production the way it performs on benchmarks. Poolside's own results show the model trailing closed-source leaders by roughly 10 to 15 percentage points on Terminal-Bench, a gap that matters for customers deciding whether self-hosting is worth the trade-off.

Whether the company can close that distance on its next cycle, while competing against Chinese open-weight models that are also improving rapidly, will determine if the Western open-weight gap is a temporary condition or a structural one.