inclusionAI releases Ling-3.0-flash-Sante, a 124B healthcare model with 5.1B active parameters

inclusionAI has released Ling-3.0-flash-Sante, a healthcare-specialized model built on Ling-3.0-flash, with 124B total parameters and 5.1B active parameters per token. Weights are available on Hugging Face and ModelScope under the identifier inclusionAI/Ling-3.0-flash-Sante. The name comes from the French word santé, meaning "health."

Key specifications

• Parameters: 124B total, 5.1B active per token (the Hugging Face page lists the safetensors model size as 127B params, with F32 and BF16 tensor types)
• Architecture: hybrid-linear MoE combining Kimi Delta Attention (KDA), gated MLA, and sparse MoE layers
• Context: training extends to 256K tokens; the reference SGLang launch uses a 262144-token context length
• Reasoning: thinking mode is on by default and can be disabled per request with {"chat_template_kwargs": {"enable_thinking": false}}
• Tool use: tool calling is supported through compatible serving frameworks
• Pricing: not yet disclosed. The model is not deployed by any Hugging Face Inference Provider at the time of reading.
• Training cutoff: not stated

Intended scope

inclusionAI says Sante targets medical reasoning, healthcare research, deep research, and tool-assisted workflows while retaining the general capabilities of Ling-3.0-flash. Listed uses include medical question answering with human review, literature search and evidence synthesis, long-document analysis, structured information extraction, healthcare knowledge-base and RAG applications, and multi-agent healthcare workflows.

Evaluation

The company claims strong results on professional healthcare tasks, complex medical reasoning, medical knowledge, and healthcare safety benchmarks. Results are presented in an evaluation figure on the model card. The benchmarks and settings named are:

• AFUMED-Drug: medication safety, measured as accuracy on medication contraindications
• AFUSAFE-MedSCE: overall safety standard of healthcare responses
• MedMCQA: standard validation set
• BrowseComp: 256K context, up to 1,000 turns. The single-agent setting uses a discard-all context-management strategy triggered at 40% context usage. The multi-agent setting uses an in-house harness that can dispatch up to 64 sub-agents per task.
• HLE with tools and DeepSearchQA: 256K context, no context-management strategy
• Tau3-Banking: official Tau evaluation set, with access restricted to the terminal_use tool

Unless otherwise specified, evaluation used temperature=1.0 and top_p=0.95, with reasoning effort set to the highest tier for all models in the comparison.

Deployment

The model card gives quickstarts for SGLang and vLLM. The SGLang example serves a 256K context on four 141GB-class GPUs or a compatible four-GPU Blackwell node, using --tp 4 and the NEXTN speculative algorithm. On 80GB GPUs such as H100 or H800, it calls for --tp 8. The vLLM example uses tensor parallel size 4, the ling3 tool-call and reasoning parsers, and MTP speculative decoding with 3 speculative tokens, which the card recommends for lower latency. Two quantizations of the model are listed on Hugging Face.

Limitations

inclusionAI states that Sante is a research model, not a medical device. Its outputs may be incomplete or incorrect, and it must not independently diagnose, prescribe, triage, or make treatment decisions. The company says high-stakes use requires qualified human review, independent verification, and compliance with privacy, security, and healthcare regulations. It also notes that benchmark results do not guarantee real-world clinical performance, and that performance can vary across languages, specialties, populations, and institutions.

What this means

Sante is a domain-tuned variant of an existing model family, released with open weights and serving recipes for two common inference stacks. Its 5.1B active parameters per token describe compute per token, but the quickstart still specifies four 141GB-class GPUs for the full 256K context, so memory requirements remain substantial. Benchmark claims come from inclusionAI's own evaluation, and the model card itself says they do not establish clinical performance. For teams building healthcare research, retrieval, or agent workflows, the stated scope is human-reviewed use, not autonomous clinical decisions.