inclusionAI: Ling 3.0 Flash Sante

Ling 3.0 Flash Sante is InclusionAI's (Ant Group) health and medicine domain fine-tune of Ling 3.0 Flash, distributed primarily through hosted APIs (for example OpenRouter, Vercel AI Gateway, and Novita) rather than as open weights on Hugging Face. It keeps the base model's hybrid-linear Mixture-of-Experts design: 124 billion total parameters with about 5.1 billion activated per token, alternating Kimi Delta Attention (KDA) with gated Multi-Head Latent Attention (MLA), plus sparse routing across hundreds of experts. The model targets a 262K-token context window, supports function calling and hybrid reasoning modes, and is positioned for medical knowledge reasoning, clinical safety, evidence-based retrieval, literature synthesis, and multi-step healthcare workflows while retaining general reasoning, coding, and agentic skills from Ling 3.0 Flash.

InclusionAI describes Sante as leading among flash-scale models on proprietary medical evaluations (vendor-reported scores on benchmarks such as MedXpertQA-Text, DiagnosisArena-MCQ, and AFUMED-Drug, plus safety-oriented sets like MedEthicAlign). Those medical suites are not part of the standard public leaderboard set used here, and as of release there is no independent third-party benchmark page (for example Artificial Analysis) dedicated to this exact fine-tune. Developers should treat Sante-specific performance claims as vendor-reported until independently verified.

Practical use cases include clinical decision-support prototypes, medical Q&A, chart and literature summarization, and tool-augmented research pipelines where a qualified professional reviews outputs. It is not intended as an autonomous diagnostic system or a substitute for licensed medical advice. For reproducible general-capability numbers on the shared Ling 3.0 Flash architecture, refer to the open base model's Hugging Face evaluation tables and Artificial Analysis listings; Sante should be selected when healthcare-focused behavior and safety alignment matter more than running the base checkpoint locally.