Poolside opens Laguna S 2.1 weights for coding agents

Poolside released Laguna S 2.1, a 118-billion-parameter Mixture-of-Experts (MoE) coding model that activates only 8 billion parameters per token and supports a context window of up to 1 million tokens. The weights are available on Hugging Face under the permissive OpenMDW-1.1 license. Poolside says the model scores 70.2% on Terminal-Bench 2.1, 78.5% on SWE-Bench Multilingual, and 59.4% on SWE-Bench Pro’s public dataset.

The release targets growing enterprise demand for open-weight systems that can run inside government, defense, and regulated environments. Poolside is positioning the model as a Western alternative to widely adopted Chinese open-weight systems, with a focus on self-hosting, lower inference costs, and faster iteration rather than frontier-scale training budgets.

Poolside also published the complete, unedited trajectory of every final benchmark trial, including reasoning steps, tool calls, and shell commands. The company disclosed reward-hacking issues seen during training and outlined mitigations such as prompt addenda, LLM-based judging calibrated against human labels, and expert review. Reported limitations include tool-schema brittleness, nested JSON errors, overthinking on competition math, and a large cost gap between thinking and non-thinking modes.