124B total / 5.1B active (MoE)
262,144 tokens
MIT
MIT-licensed healthcare-specialized version of Ling-3.0-flash designed for medical reasoning, healthcare research, evidence synthesis, deep research, tool-assisted workflows, and multi-agent collaboration. It retains the BailingMoeV3 hybrid-linear Mixture-of-Experts architecture with 42 transformer layers, including 35 Kimi Delta Attention layers and seven gated Multi-head Latent Attention layers, two initial dense layers, 512 routed experts plus one shared expert, and eight routed experts activated per token. The architecture uses a 2,560-dimensional hidden state, 32 attention and key/value heads, a 157,184-token vocabulary, and one NEXTN multi-token-prediction layer for speculative decoding. inclusionAI designates the model as 124B total parameters with 5.1B active per token, while the released safetensors index inventories approximately 127.5B serialized parameters across 27 BF16 shards totaling roughly 255 GB. The model supports 262,144-token context, thinking and non-thinking operation with thinking enabled by default, structured tool calling, and deployment through SGLang or vLLM with MTP speculative decoding. Publisher-reported healthcare results include 54.3 on HealthBench Professional, 79.2 on DiagnosisArena-MCQ, 89.7 on ReDisQA, 51.0 on MedXpertQA-Text, 94.2 on MedQA-USMLE, 80.3 on MedMCQA, and 90.6 on an internal drug benchmark. Reported research and agent results include 80.2 on BrowseComp with single-agent context, 90.8 on DeepSearchQA, 57.1 on Humanity's Last Exam with tools, 44.3 on Tau3-Banking, 75.6 on MCP Atlas, and 70.5 on AA-LCR. Official deployment examples use four 141 GB-class GPUs or eight 80 GB H100/H800 GPUs. Sante is explicitly a research model rather than a medical device and must not independently diagnose, prescribe, triage, or make treatment decisions; healthcare outputs require review by qualified professionals.