Poolside releases new coding model, claims performance matching larger rivals
AI lab Poolside has released a new open coding model, Laguna S 2.1. The company claims the model performs on par with, or better than, competitors ten times its size on certain tasks.
San Francisco-based AI lab Poolside released its latest coding model, Laguna S 2.1, on Tuesday. The model is a 118-billion-parameter Mixture-of-Experts (MoE) system that activates only 8 billion parameters per token and supports a context window up to 1 million tokens. Poolside asserts that the model's performance in benchmarks matches or exceeds that of significantly larger open models competing for similar use cases.[newline]
According to Poolside, Laguna S 2.1 achieved 70.2% on the Terminal-Bench 2.1 benchmark, placing it higher than several larger models, including DeepSeek-V4-Pro-Max (1.6 trillion parameters) and Nvidia's Nemotron 3 Ultra (550 billion parameters). The development from pre-training to public release took under nine weeks, utilizing 4,096 Nvidia H200 GPUs. This rapid cycle aligns with the company's previous release of three models in three months.[newline]
The Laguna S 2.1 release comes amid a contentious debate surrounding open-source AI models. Adoption has recently shifted towards open systems that companies can download and run on their own infrastructure, with many leading open models originating from Chinese labs. Poolside positions Laguna S 2.1 as a response to this need for Western entities, aiming to provide open models that can be trusted, run, and built upon, as stated by co-CEO Jason Warner.[newline]
The model employs a sparse MoE architecture, meaning inference costs scale with active parameters (8 billion) rather than total parameters (118 billion). Poolside highlights that the model is small enough to operate on a single Nvidia DGX Spark. In terms of pricing, the company offers aggressively competitive access to the model, undercutting most frontier alternatives by a significant margin.[newline]
Poolside also aims to increase transparency in AI model evaluation by publishing all benchmark trial trajectories unedited. This, combined with the model's performance and competitive pricing, seeks to solidify its position among enterprise clients requiring reliable and high-performing AI solutions within their own security boundaries.