New model releases: ultra-sparse MoE, metabolic research and scientific AI

An open-weight multimodal ultra-sparse mixture-of-experts model was released on August 26, aimed at coding, office, text and vision tasks. It has 125 billion main parameters plus 51 billion N-gram embeddings, with 6 billion active per token and a 262,000-token open context; a managed version offers a one-million-token context and tools. Limitations include a large memory footprint, vendor-reported benchmarks and deployment complexity. Weights are available; managed pricing had not been independently verified at the cutoff.

A self-supervised continuous-glucose-monitor foundation model for metabolic prediction research was also released, using slow-state and event dual streams with few-shot transfer. It is a preprint with author-reported metrics and no clinical approval.

A neural-operator scientific AI model with an enterprise focus was announced, but public weights, pricing and independent benchmarks are not yet available.