West Responds to China: Poolside Releases Most Powerful Open Model Laguna S 2.1

Startup Poolside has released Laguna S 2.1, an open model for agentic programming with 118 billion parameters, calling it the most powerful open model in the West. It uses a Mixture-of-Experts architecture: out of 118 billion parameters, only 8 billion are active per token, with a context window of up to one million tokens. On benchmarks, the model scores 70.2% on Terminal-Bench 2.1 and 59.4% on SWE-Bench Pro. This is on par with or above DeepSeek V4 lite, NVIDIA's Nemotron 3 Ultra, and Thinking Machines' Inkling, even though those models are 2.5 to 8 times larger. The company has published the full trajectory of how Laguna independently proved Erdős problem #397, as well as trajectories from all benchmark runs for reproducibility. Over the past 11 months, Western labs have not released open models of this scale since OpenAI's gpt-oss-120b. Poolside claims to have built a 'model factory' that automates experiments, training Laguna S 2.1 in 4 weeks on 4,000 H200 GPUs, with releases promised every 5 weeks. However, amidst Chinese competition, Laguna lags behind Moonshot AI's Kimi K3: K3 scores 88.3% on Terminal-Bench 2.1 with 2.8 trillion parameters versus Laguna's 118 billion. The release comes amid discussions in the US about possibly banning the use of Chinese open-source models. Moonshot promises to release Kimi K3's weights on July 27.