Ling 3.0 Flash Sante Guide: A Free Medical Model, Unbenchmarked

A free medical model is an easy thing to get excited about and an easy thing to misuse. Ling 3.0 Flash Sante is both: a health-focused variant of Ling 3.0 Flash that OpenRouter listed with a no-cost endpoint at launch, with almost nothing published about how well it performs.

That second fact should shape how you use it. This guide sets out what is actually known, what is not, and a way to use the model that does not depend on the missing information.

What is actually known

Sante is built on Ling 3.0 Flash, so it inherits the base architecture: a 124B mixture-of-experts model with about 5.1B active parameters per token, a 262,144-token context window, text in and text out, reasoning mode and tool calling. OpenRouter listed it from 4 September 2026 as a :free endpoint, though it no longer appeared in OpenRouter's public model list at the time of writing.

The stated focus, from the provider listings, is medical knowledge reasoning, clinical safety, evidence-based retrieval and long-horizon medical tasks, while keeping the base model's general reasoning, coding and agent abilities. Novita, which also hosts it, lists the same description with a paid price rather than a free one.

What is missing

Three things you would normally check before adopting a model are not available.

• Benchmarks. We found no published scores, from inclusionAI or anyone else, and the provider listings make qualitative claims only. Nothing public supports a clinical-accuracy ranking.
• Training method. The base and Fin models come with model cards that describe their training. We could not find an equivalent card for Sante.
• Public weights. The base model and the Fin variant are on Hugging Face under the MIT licence. inclusionAI's Hugging Face listing showed no Sante weights at the time of writing, so assume it is API-only until inclusionAI says otherwise.

Missing weights matter more in medicine than elsewhere. Many health organisations cannot send patient data to a third-party API at all, and for them a model they cannot self-host is a model they cannot use for anything involving real records.

A useful comparison point

There is one indirect signal. The model card for Ling 3.1 Flash, inclusionAI's larger general model, lists HealthBench Professional among its evaluations and names healthcare alongside STEM and finance as areas where it performs strongly. So inclusionAI is now pursuing medical ability in two ways: a small specialist and a large generalist.

If you are building something health-related, test both. A specialist fine-tune of a 5.1B-active model may know the vocabulary better, but a 25B-active generalist may simply know more. Which effect wins on your task is an empirical question, and with no public numbers for Sante it is one only you can answer.

Where it is reasonable to use

The safe pattern for any medical model, and especially an unbenchmarked one, is to keep it on tasks where a mistake is caught before it matters:

• Summarising published literature for a clinician who will read the sources.
• Drafting plain-language explanations that a qualified person reviews before release.
• Classifying or routing documents where a wrong label sends something to the wrong queue rather than to a patient.
• Writing and testing code for health software, where medical vocabulary in comments and data models helps.

Where it is not reasonable: diagnosis, treatment recommendations, anything a patient reads unreviewed, and any tool that can change a record. No model listing positions it as a medical device or a substitute for clinical judgement, and you should not either. Keep tool permissions narrow. PII in prompts covers how patient data leaks into requests in the first place, which is the problem to solve before any experiment with real data.

The free endpoint is for evaluation

The :free variant is the same model on a rate-limited, no-cost endpoint. That is ideal for building a test set and running it, and unsuitable for production: free endpoints throttle, can disappear, and promotional pricing has an end date. This one already illustrates the point: Novita charges for the model, and the OpenRouter free listing had gone by the time of writing. Free AI API tiers compared explains how to plan around limits that move.

A minimum evaluation before you rely on it

1. Write thirty questions from your own domain with answers a clinician has signed off.
2. Include ten where the correct answer is "this needs a clinician" or "the evidence is unclear". Safety shows up as refusing to overreach.
3. Run Sante, base Ling 3.0 Flash and Ling 3.1 Flash on the same set with the same retrieval.
4. Score separately for factual errors, fabricated citations and overconfident answers.

If Sante does not clearly beat the base model on your set, the specialist label is not earning anything, and the base model has the advantage of published weights.