Poolside has introduced Laguna S 2.1, a 118-billion-parameter open-weight model specifically built for agentic coding. The San Francisco-based startup claims that this model matches or even surpasses models several times its size in performance. The model employs a mixture-of-experts (MoE) architecture, activating eight billion parameters per token, making it efficient enough to run on a single Nvidia DGX Spark desktop system. The weights are freely available on Hugging Face under the Linux Foundation's OpenMDW license.
On benchmark evaluations such as Terminal-Bench and SWE-Bench Pro, which measure agentic coding capabilities, Laguna S 2.1 achieved scores just over 70 percent and nearly 60 percent, respectively. These results are comparable to or better than models from DeepSeek, Nvidia, and Thinking Machines that carry two to eight times as many active parameters. However, Poolside acknowledges that the model is “not yet at the frontier,” as closed-source systems from OpenAI and Anthropic still score significantly higher on the same evaluations.
This release is framed as a direct response to the dominance of Chinese laboratories in the open-weight model category. For over a year, organizations like DeepSeek, Alibaba’s Qwen family, and Moonshot’s Kimi have set the pace in open-weight AI development. No Western lab had released a model in the 118-billion-parameter class for 11 months prior to this launch. Forbes reported that Poolside explicitly positioned the release as an effort to provide Western enterprises and governments with a self-hosted alternative that can be run without sending data to a foreign provider.
The history of open-weight models dates back to early initiatives like BERT and GPT-2, which were open-sourced to encourage research and development. However, the landscape shifted with the emergence of DeepSeek’s R1 and Alibaba’s Qwen2.5, which demonstrated that open-weight models could achieve cutting-edge performance. Western companies began to feel pressure, as many enterprises preferred open-weight models for data privacy and customization. Poolside’s Laguna S 2.1 aims to fill this gap, offering a robust alternative that can be deployed on-premises or in secure environments.
In terms of technical details, the mixture-of-experts architecture allows Laguna S 2.1 to allocate computational resources efficiently. Unlike dense models where all parameters are active for every token, MoE models activate only a subset of parameters, reducing inference costs and enabling deployment on less powerful hardware. This design choice is crucial for enterprise adoption, as it lowers the barrier to entry for self-hosting. The model was trained using Poolside’s internal Model Factory platform, which automates architecture search and reinforcement learning from code execution. Training was completed in under four weeks using 4,000 Nvidia H200 GPUs.
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 funding round in October 2024, achieving 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 withdrew from a joint data centre project in Texas. Despite this setback, Poolside continues to serve government, defense, and other highly regulated organizations through its API and agent harness.
The smaller Laguna XS model was 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. This showcases the model’s ability to handle complex tasks, which is critical for agentic coding applications where autonomous agents must execute multi-step workflows.
The bet behind Laguna S 2.1 is that enterprises will prefer to pay for a capable coding model that runs on their own hardware rather than sending prompts to a closed API. This thesis depends on the model performing in production as well as it does on benchmarks. Poolside’s own results show that Laguna S 2.1 trails closed-source leaders by approximately 10 to 15 percentage points on Terminal-Bench. For customers deciding whether self-hosting is worth the trade-off, this gap may be significant. However, the benefits of data sovereignty, customization, and potential cost savings could tip the scale if the model improves on subsequent releases.
The Western open-weight gap is not just a technical issue; it also has geopolitical implications. With Chinese labs dominating the open-source AI landscape, Western governments are concerned about dependence on foreign technology for critical infrastructure. Poolside’s model offers a self-hosted solution that can be audited, modified, and deployed without reliance on Chinese supply chains. This aligns with broader efforts to ensure digital sovereignty and security.
From a technical perspective, the mixture-of-experts architecture used in Laguna S 2.1 is gaining traction across the industry. Models like Mixtral 8x7B by Mistral and DBRX by Databricks have shown that MoE can achieve high performance with fewer active parameters. Poolside’s approach leverages a larger total parameter count (118 billion) but keeps the active set small, aiming to balance quality and efficiency. The OpenMDW license used by Poolside is permissive, encouraging broad adoption and community contributions.
The benchmarks on Terminal-Bench and SWE-Bench Pro are particularly relevant for agentic coding. These evaluations test a model’s ability to interact with a terminal, execute commands, and debug code autonomously. Achieving 70% on Terminal-Bench means that Laguna S 2.1 can handle a substantial portion of common software engineering tasks without human intervention. This capability is valuable for companies looking to automate parts of their development pipelines.
Poolside’s competitors in the Western open-weight space include Code Llama from Meta, StarCoder from Hugging Face and ServiceNow, and WizardCoder from Microsoft. However, these models typically have fewer parameters and may not offer the same level of agentic capability. Code Llama 70B, for instance, is a dense model that does not use MoE, making it less efficient for deployment at scale. Moreover, none of these models are specifically built for agentic workflows, which involve multi-step reasoning and tool use.
The company’s rapid release cycle suggests an aggressive development timeline. With a new model shipped every five weeks, Poolside aims to iterate quickly and close the gap with frontier models. The use of reinforcement learning from code execution (RLCE) is a novel training technique that reinforces correct code generation by rewarding successful runs. This method goes beyond static training data and teaches the model to reason about the outcome of its code.
While Laguna S 2.1 may not yet match the frontier performance of GPT-4 or Claude, it represents a significant step toward democratizing advanced coding AI. For enterprises and governments that prioritize data privacy and control, the trade-off of a slight performance hit may be acceptable. Poolside’s challenge will be to demonstrate that its model can consistently deliver high-quality results in real-world scenarios, not just on curated benchmarks.
As the open-weight AI landscape evolves, the competition between Western and Chinese labs will drive innovation. Poolside’s entry into the 118B-parameter class is a welcome addition, but whether it can maintain its pace and close the performance gap will determine its long-term impact. The company’s focus on enterprise and government customers provides a clear market niche, and its ability to serve these clients without compromising performance will be tested in the coming months.