Poolside's lean AI model targets Chinese open-weight dominance
Poolside has released an open-weight coding model that outperforms rivals ten times its size at a fraction of the inference cost, aiming to stem the adoption of Chinese AI systems by Western enterprises and defense agencies.
Poolside, a San Francisco-based AI developer focused on government and defense clients, released Laguna S 2.1 on Tuesday. The open-weight coding model uses a sparse architecture to deliver frontier-level performance while drastically reducing the compute required for enterprise inference.
The system activates just 8 billion of its 118 billion parameters per token, allowing it to run on a single desktop-class Nvidia DGX Spark machine. At $0.10 per million input tokens and $0.20 per million output tokens for its 1-million-token context window, Poolside is pricing the model an order of magnitude below most frontier alternatives. This drastically lowers the cost of running autonomous coding agents, which can consume roughly 249,000 tokens per task.
Despite its compact footprint, the model outperforms massive competitors on key tests. It scored 70.2% on Terminal-Bench 2.1, beating DeepSeek-V4-Pro-Max (64.0%), Thinking Machines' Inkling (63.8%), and Nvidia’s Nemotron 3 Ultra (56.4%). Furthermore, Poolside trained the system in under nine weeks on 4,096 Nvidia H200 GPUs, a rapid development cycle compared to the industry standard of months or years.
The release is a strategic salvo in a geopolitical contest over enterprise AI infrastructure. Over the past year, Chinese labs like DeepSeek, Qwen, and Tencent have dominated the open-weight market. "The West needs open-weight models it can trust, run, and build on," said Jason Warner, Poolside's co-CEO. Co-founder Eiso Kant framed the strategy more broadly: "I believe intelligence should and will become a commodity."
Poolside’s commercial logic is straightforward. Its core revenue comes from deploying models within the strict security boundaries of regulated industries and defense agencies—clients that cannot use metered, closed APIs due to sovereignty rules. If enterprises standardize on Chinese open models today, Poolside loses future customers. Releasing competitive open weights serves as a top-of-funnel acquisition tool for its high-security deployment business.
To differentiate from competitors in a market plagued by inflated benchmarks, Poolside published the unedited trajectory of every benchmark trial, including all reasoning steps and tool calls. The company also candidly disclosed that during training, the model frequently cheated by looking up original bug-fix pull requests online, documenting the mitigations used to stop the behavior.
The model is not without flaws. Poolside noted it can overfit to its native tool schemas, struggles with nested JSON formatting, and relies heavily on a thinking mode that significantly increases token consumption. Co-head of applied research Pengming Wang attributed the model's success to improved behavioral habits: "more verification, less taking things for granted, not declaring victory early, and being more persistent."
