What is Ling-3.0-flash-Sante?
Ling-3.0-flash-Sante is a healthcare-specialised language model from inclusionAI, Ant Group's AI research arm, published on Hugging Face on 19 October 2026. Its name is the French santé, health. It is built on Ling-3.0-flash, not the larger Ling-3.1-flash: a hybrid-linear mixture-of-experts model with 124 billion parameters, 5.1 billion active per token, combining Kimi Delta Attention, gated MLA and sparse expert layers.
It is trained for medical reasoning, healthcare research, deep research and tool use, and keeps the general skills of its base. Training extends to a 256K-token context, thinking is on by default and can be switched off per request, and tool calling works through SGLang and vLLM. inclusionAI calls it a research model, not a medical device, and says it must not diagnose, prescribe or triage on its own.
The weights are under the MIT licence. Vercel's AI Gateway serves it through one provider, Novita, at $0.07 per million input tokens and $0.22 per million output tokens on its headline price, with Novita's row showing $0.08 input.
Is Ling-3.0-flash-Sante worth it?
Ling-3.0-flash-Sante makes a narrow, well-supported pitch: near-frontier medical answers from a model that activates only 5.1 billion parameters. On inclusionAI's own table it scores 54.3 on HealthBench Professional, ahead of Claude Opus 4.8 (49.4), Kimi K3 (49.6) and GLM-5.3-Flash (49.1) and behind only GPT-5.6 Sol (60.8). On MedQA-USMLE it reaches 94.2 against Sol's 96.7, and 90.6 on Ant's internal drug-contraindication test. These are self-reported, published only as an image, and three of the medical rows are Ant's own internal benchmarks; we found no independent evaluation.
The weaknesses are in the same table. On MedXpertQA, a harder exam set, it scores 51.0 against 62.4 for Gemini 3.6 Flash; on MedMCQA it is last among the large models at 80.3; and on AA-LCR long-context reasoning its 70.5 trails every model listed except Nemotron-3-Super and Step-3.7-Flash. Its deep-research scores are strong for its size, 80.2 on BrowseComp with Ant's context management, but those depend on Ant's harness settings.
It is distinct from Ling-3.1-flash, which is a 560B general model. Sante is a quarter of that size and built on the earlier 3.0 base, so it is cheaper to host: the card's reference setup is four 141 GB GPUs, or eight 80 GB H100s. Through Vercel it costs $0.07 input and $0.22 output per million tokens, about three and a half times the base Ling-3.0-flash's $0.02 input on the same gateway but far below frontier prices.
For health data the important point is the MIT licence. A hospital or health-tech team can run it inside its own network with no usage restrictions, which matters more for patient records than the last few benchmark points. The hosted route goes through Novita, whose zero-retention and training terms Vercel's page does not mark.
Where it earns its place
An ML lead at a 300-bed hospital's informatics department
Summarise trial abstracts and compare study protocols without patient data leaving the network
Serve the MIT-licensed weights with SGLang on four H200-class GPUs at a 256K context, behind clinician review. The licence puts no conditions on commercial or clinical-research use; regulatory compliance stays with the hospital.
A product manager at a health-information startup
Answer consumer medication questions at low cost, with a pharmacist checking flagged replies
Through Vercel at $0.07 input and $0.22 output per million tokens, a million 1,000-token questions with 500-token answers costs about $180 before reasoning tokens. Ant's 90.6 on its own drug-contraindication test is encouraging but unverified.
Before you commit
- Setup
- Pull the Ling-3.0 SGLang image and launch with tensor parallelism 4 on 141 GB-class GPUs (8 on H100 or H800), or build current vLLM with the ling3 tool and reasoning parsers. Hosted, call inclusionai/ling-3.0-flash-sante on Vercel AI Gateway.
- Integrations
- OpenAI-compatible servers through SGLang and vLLM with tool calling and MTP speculative decoding; Vercel AI Gateway exposes Chat Completions, Responses and Anthropic Messages formats.
- Data and privacy
- MIT weights allow unrestricted self-hosting, the safer route for patient data. On Vercel the only provider is Novita, and the gateway page does not mark it as zero-retention or no-training; check Novita's terms before sending health information.
- Cost at scale
- Self-hosting needs four 141 GB GPUs or eight 80 GB GPUs for the full 256K context. Hosted, 100 million input and 20 million output tokens a month is about $11.40 at Vercel's headline price, before reasoning tokens.
- Limits
- 256K-token context and up to 32,000 output tokens on Vercel, reasoning included. Text only. inclusionAI warns it can produce outdated or unsafe medical information and must not diagnose, prescribe or triage independently.
What’s good
- 54.3 on HealthBench Professional, above Opus 4.8 (self-reported)
- MIT licence; self-hostable on four 141 GB GPUs
- $0.07 / $0.22 per 1M tokens on Vercel AI Gateway
- 256K context, tool calling, thinking on or off
What’s not
- No independent evaluation; results published only as an image
- Trails on MedXpertQA, MedMCQA and long-context reasoning
- One hosted provider, without stated zero-retention terms
- Research model; not for unreviewed clinical decisions
Best for
Health-tech and research teams that want an open, permissively licensed medical model to self-host behind human review.
Ling-3.0-flash-Sante alternatives
Ling-3.0-flash-Sante: common questions
- Is Ling-3.0-flash-Sante worth it?
- A cheap, MIT-licensed medical specialist that beats Claude Opus 4.8 on Ant's own HealthBench Professional run, though only Ant has measured it and it still trails GPT-5.6 Sol. We rate it 3.8 out of 5.
- How much does Ling-3.0-flash-Sante cost?
- Open weights under MIT. Vercel AI Gateway: $0.07 input / $0.22 output per 1M tokens (Novita's row: $0.08 input, $0.02 cached).
- What are the downsides of Ling-3.0-flash-Sante?
- No independent evaluation; results published only as an image; trails on MedXpertQA, MedMCQA and long-context reasoning; one hosted provider, without stated zero-retention terms; research model; not for unreviewed clinical decisions.
- Who is Ling-3.0-flash-Sante best for?
- Health-tech and research teams that want an open, permissively licensed medical model to self-host behind human review.
- How much does Ling-3.0-flash-Sante cost for a team or in production?
- Self-hosting needs four 141 GB GPUs or eight 80 GB GPUs for the full 256K context. Hosted, 100 million input and 20 million output tokens a month is about $11.40 at Vercel's headline price, before reasoning tokens.
- Is Ling-3.0-flash-Sante safe for company data?
- MIT weights allow unrestricted self-hosting, the safer route for patient data. On Vercel the only provider is Novita, and the gateway page does not mark it as zero-retention or no-training; check Novita's terms before sending health information.
- What are the best Ling-3.0-flash-Sante alternatives?
- The closest alternatives we have reviewed are Ling-3.1-flash, GLM-5.3-Flash, Kimi K3 and DeepSeek-V4.1-Flash.
Answers checked Oct 31, 2026.