Four teams in the government-backed \"sovereign AI foundation model\" project have completed preparations for a second-round review, industry sources said. Two large companies have built models with more than 600 billion total parameters, comparable in size to major global open-weight models. Two startups opted for models of around 300 billion parameters, focusing on compute efficiency and task execution. All four use a mixture-of-experts structure and are pursuing different approaches to scaling up.

AI coding startup Poolside said it has unveiled an open-weight foundation model for coding agents called Laguna S 2.1. The company said the model has 118 billion parameters and showed performance that matched or exceeded several larger models on Terminal-Bench 2.1 and SWE-Bench Pro. It said the model can run on a single Nvidia DGX Spark system and that its weights are available on Hugging Face under the OpenMDW-1.1 license.