Poolside has officially launched Laguna S 2.1, a 118-billion-parameter open-weight model tailored for agentic coding tasks. The San Francisco-based startup claims this new model matches or even surpasses the performance of competing models that are two to eight times larger in terms of active parameters. Built on a mixture-of-experts (MoE) architecture, Laguna S 2.1 activates only eight billion parameters per token, making it remarkably efficient. The model is compact enough to run on a single Nvidia DGX Spark desktop system, a significant advantage for organizations that require on-premise deployment. The weights are publicly available on Hugging Face under the Linux Foundation’s OpenMDW license, ensuring broad accessibility and transparency.
On two key benchmarks for agentic coding—Terminal-Bench and SWE-Bench Pro—Laguna S 2.1 achieved scores of just over 70% and nearly 60% respectively. These results are competitive with models from DeepSeek, Nvidia, and Thinking Machines that have far larger active parameter counts. However, Poolside acknowledges that the model is “not yet at the frontier,” noting that closed-source systems from OpenAI and Anthropic still outperform it by a margin of roughly 10 to 15 percentage points on Terminal-Bench. This gap is critical for customers weighing the trade-offs between self-hosting and using a closed API.
Strategic Context: The Western Open-Weight Gap
The release of Laguna S 2.1 is framed as a direct response to the dominance of Chinese labs in the open-weight category. For more than a year, models from DeepSeek, Alibaba’s Qwen family, and Moonshot’s Kimi have set the pace in open-source AI development. According to the company, no Western lab had released an open-weight model in the 118-billion-parameter class for 11 months before this launch. 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. This is particularly relevant for sectors like defense, government, and regulated industries where data sovereignty is paramount.
The lack of competitive Western open-weight models has been a growing concern for policymakers and industry leaders. The United States and Europe have invested heavily in AI research but have primarily focused on closed-source or API-based models. Chinese labs, on the other hand, have aggressively released open-weight models that rival the best in the world, often with lower costs and broader accessibility. Poolside’s Laguna S 2.1 aims to close that gap by providing a Western-made model that can be self-hosted, customized, and audited. The company’s pitch is that enterprises will prefer to run a capable coding model on their own hardware rather than relying on a foreign API, especially when dealing with sensitive codebases or proprietary algorithms.
Company Background and Funding
Poolside was founded in 2023 by Jason Warner, former chief technology officer at GitHub, and Eiso Kant. The company raised $500 million in a Series B round in October 2024, achieving a valuation of $3 billion 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 center project in Texas. Despite this setback, Poolside continues to serve government, defense, and other highly regulated organizations through its API and agent harness. The company’s focus on agentic coding—where models not only generate code but also execute and debug it autonomously—differentiates it from general-purpose LLMs.
Technical Details and Training
Poolside developed Laguna S 2.1 using its internal Model Factory platform, which automates architecture search and reinforcement learning from code execution. The training process was completed in under four weeks on 4,000 Nvidia H200 GPUs. The smaller Laguna XS model was launched three weeks earlier, and the company says it ships new models on roughly a five-week cadence, suggesting a rapid iteration cycle. 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. This showcases the model’s ability to handle complex, multi-step tasks that require planning and logical deduction.
The MoE architecture is a key factor in Laguna S 2.1’s efficiency. By activating only a subset of parameters per token, the model reduces computational costs while maintaining high performance. This makes it feasible to run on a single DGX Spark, which is a compact desktop system designed for AI workloads. The ability to deploy such a powerful model on local hardware is a compelling proposition for enterprises that want to avoid the latency, cost, and privacy concerns associated with cloud-based APIs.
Competitive Landscape and Benchmark Performance
While Laguna S 2.1 performs admirably on agentic coding benchmarks, it still lags behind the best closed-source models. On Terminal-Bench, Poolside’s own results show a gap of roughly 10 to 15 percentage points compared to OpenAI’s GPT-4 and Anthropic’s Claude models. This gap matters for customers deciding whether self-hosting is worth the trade-off. However, for many use cases, the model’s performance may be sufficient, especially when combined with the benefits of data privacy and customization. The company argues that the benchmark gap is closing with each release and that the rapid iteration cycle will soon bring its models to the frontier.
The Chinese open-weight models, such as DeepSeek-Coder and Qwen2.5-Coder, are also improving rapidly. DeepSeek recently released a model that achieved state-of-the-art results on several coding benchmarks, and Alibaba’s Qwen family continues to expand with specialized variants. This creates a dynamic where the “Western open-weight gap” could be either a temporary condition or a structural one, depending on how quickly Poolside and other Western labs can innovate. Poolside’s focus on agentic coding might give it an edge in certain domains, but the overall competition is fierce.
Implications for Enterprises and Governments
The release of Laguna S 2.1 comes at a time when many organizations are reconsidering their AI strategies. The push for data sovereignty, coupled with increasing regulatory scrutiny around AI usage, has made self-hosted models more attractive. For enterprises in defense, finance, healthcare, and critical infrastructure, the ability to run an AI model on-premises without sending data to third parties is a significant advantage. Poolside’s model, with its open-weight license and efficient architecture, offers a credible alternative to both closed APIs and Chinese open-weight models.
However, the model’s performance gap means that enterprises must carefully evaluate their needs. If they require state-of-the-art performance for complex coding tasks, they may still need to rely on closed APIs from US-based providers like OpenAI or Anthropic. But for many routine coding tasks—such as code completion, bug fixing, and unit test generation—Laguna S 2.1 may be more than adequate. The decision will also depend on the cost of running the model locally versus paying per-token for an API service.
Poolside is also positioning itself as a partner for highly regulated organizations through its agent harness, which allows models to be integrated into existing workflows with safeguards. The company’s experience serving government and defense clients suggests it understands the compliance and security requirements of these sectors. Whether the model can deliver in production environments remains to be seen, but the benchmarks are promising.
Looking ahead, Poolside’s five-week release cycle means we can expect new models frequently. The company’s bet is that enterprises will pay to run a capable coding model on their own hardware rather than send prompts to a closed API. This thesis depends on Laguna S 2.1 performing in production the way it performs on benchmarks. If Poolside can close the gap with frontier models in the next few iterations, it could become a major player in the agentic coding space. If not, it may remain a niche option for organizations that prioritize data control over raw performance.
In the broader context, the Western open-weight gap is a reflection of the larger geopolitical dynamics in AI. Chinese labs have benefited from government support, a large talent pool, and a willingness to release models as open source. Western labs, led by companies like Meta with its Llama series and now Poolside, are trying to counter that by offering competitive open-weight models. The success of Laguna S 2.1 will be an important indicator of whether the West can keep pace in the open-source AI race.