Poolside, a San Francisco AI lab that has spent most of its three-year existence selling coding models to governments and defense agencies, released Laguna S 2.1 on Tuesday, July 21, 2026. The 118-billion-parameter Mixture-of-Experts (MoE) model activates only 8 billion parameters per token and supports a context window of up to 1 million tokens. Weights went live immediately on Hugging Face under the permissive OpenMDW-1.1 license.<\/p>\n
Laguna S 2.1 is an open-weight coding model from Poolside built around a sparse MoE architecture. The model uses 256 routed experts plus one shared expert, grouped-query attention, and interleaved sliding-window layers. Because inference cost scales with the 8 billion active parameters rather than the full 118 billion, Poolside says the model is small enough to run on a single Nvidia DGX Spark.<\/p>\n
On Terminal-Bench 2.1, which measures long-horizon terminal tasks, Laguna S 2.1 scores 70.2 percent, placing 11th on the company’s compiled leaderboard. It runs ahead of DeepSeek-V4-Pro-Max, a 1.6-trillion-parameter model that scored 64.0, Thinking Machines Inkling at 975 billion parameters and 63.8, and Nvidia Nemotron 3 Ultra at 550 billion parameters and 56.4.<\/p>\n
On SWE-Bench Multilingual, Laguna S 2.1 posts 78.5 percent, and on the SWE-Bench Pro public dataset it reaches 59.4 percent. With thinking mode enabled on its hardest benchmark, the model consumes roughly 249,000 completion tokens per trajectory.<\/p>\n
Pre-training began May 22, and the model launched in under nine weeks, trained on 4,096 Nvidia H200 GPUs. Poolside has now shipped three models in three months.<\/p>\n
Poolside frames the release as a response to the dominance of Chinese open-weight labs, naming DeepSeek, Qwen, Kimi, GLM, MiniMax, and Tencent Hunyuan. The company notes that Laguna S 2.1 occupies a size class into which no Western lab has released open weights in 11 months, since OpenAI’s gpt-oss-120b last August.<\/p>\n
Co-CEO Jason Warner said, “The West needs open-weight models it can trust, run, and build on.” Co-founder and co-CEO Eiso Kant wrote on X, “I believe intelligence should and will become a commodity.”<\/p>\n
Anyone can download the weights from Hugging Face under the OpenMDW-1.1 license, which the company describes as permissive. Because only 8 billion parameters activate per token, the model can run on a single Nvidia DGX Spark, lowering the hardware bar for self-hosting.<\/p>\n
Laguna S 2.1 is a 118-billion-parameter Mixture-of-Experts coding model released by Poolside on July 21, 2026. It activates 8 billion parameters per token, supports a 1 million token context window, and is available on Hugging Face under the OpenMDW-1.1 license.<\/p>\n
On Terminal-Bench 2.1 it scores 70.2 percent, ahead of DeepSeek-V4-Pro-Max at 64.0, Thinking Machines Inkling at 63.8, and Nvidia Nemotron 3 Ultra at 56.4. It also scores 78.5 percent on SWE-Bench Multilingual and 59.4 percent on SWE-Bench Pro.<\/p>\n
Poolside positions the release as a Western response to leading Chinese open-weight labs, noting that no Western lab has released open weights in this size class for 11 months, since OpenAI’s gpt-oss-120b. Co-CEO Jason Warner said the West needs open-weight models it can trust, run, and build on.<\/p>\n