InclusionAI Releases Ling-3.0-flash-Sante Healthcare MoE Model
InclusionAI has released Ling-3.0-flash-Sante, a healthcare-specialized mixture-of-experts model featuring 124B total parameters with 5.1B active parameters per token. The model weights are now available on Hugging Face and ModelScope for research and local deployment. This release provides local LLM practitioners and medical researchers with an open-weights, domain-specific model optimized for complex medical reasoning, literature research, and tool-assisted workflows. It represents a shift toward targeted domain releases rather than incremental improvements on general-purpose leaderboards. Sante utilizes a hybrid-linear mixture-of-experts architecture combining Kimi Delta Attention, gated MLA, and sparse MoE layers while supporting a 256K token context window. It comes with native reasoning and tool-calling capabilities enabled by default, though it is intended strictly as a research model rather than a medical diagnostic device.
## BACKGROUND
Mixture of Experts (MoE) is a neural network architecture where multiple parallel sub-networks, or 'experts', are used, but only a small subset of them are activated for any given token to reduce computational cost. ModelScope is an open-source model community platform offering models, datasets, and interactive demos, similar to Hugging Face.