Poolside released Laguna S 2.1, a 118-billion-parameter sparse MoE model (8B active parameters per token) under a permissive open-weight license, reporting scores of 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-Bench Multilingual, ahead of several much larger models including DeepSeek-V4-Pro-Max (1.6T parameters) and Nvidia Nemotron 3 Ultra (550B). The model was trained on 4,096 H200 GPUs over under nine weeks, priced at $0.10/$0.20 per million input/output tokens, and is self-hostable on a single DGX Spark, directly targeting government, defense, and regulated enterprise customers who cannot use closed API-gated models. Structurally, the release extends the case that efficiency gains and architectural choices (sparse MoE, post-training on agentic environments) are compressing the capability-per-compute ratio faster than raw scale alone, while also illustrating that Western open-weight supply in the competitive coding-model tier has been thin for nearly a year. For financial institutions, this is a data point on inference cost trajectories and the competitive viability of smaller, self-hosted models against both closed frontier labs and Chinese open-weight alternatives.
363 other tracked developments touching the same claims.