For a signed-in support portal, use Ling 3.0 Flash Sante only after confirming which host can serve it and whether your local agent can call that host. The model is a health-focused research model. For this browser task, limit it to non-sensitive test tickets or a synthetic queue, and ask it to return ticket IDs and page links for human review. Do not use it to diagnose, triage, prescribe, or process real patient records. InclusionAI’s Hugging Face card says it is a research model, not a medical device, and warns that outputs can be wrong. (model card)
A LocalLLaMA discussion about the Sante release included a user saying they were glad InclusionAI kept its 120B and 8B MoE models available. That is a preference about model availability, not evidence that Sante is safe for clinical decisions or ready for your environment. (Reddit discussion)
Key takeaways
Sante is the healthcare-focused Ling 3.0 Flash model, but its card labels it a research model, not a medical device.
The worktree topic row calls it closed/API only; current sources show public weights on Hugging Face and a hosted Vercel AI Gateway listing. Verify your provider before configuring a client.
Keep real health information out of the prompt. Use synthetic fixtures or non-sensitive operational records only.
Have the agent list visible ticket identifiers and source links, then stop at a human review checkpoint.
Verify the model host before starting
| Route | What the source confirms | What you still need to check |
|---|---|---|
| InclusionAI model card | Public weights and intended research uses | Runtime support, hardware, license obligations, and safe deployment |
| Vercel AI Gateway | A hosted listing for inclusionai/ling-3.0-flash-sante | Current account access, region, limits, and client SDK support |
| Local agent client | The client may support a configured provider | Whether it can run Tabbit CLI locally as well |
| Tabbit Browser | A local signed-in browser profile | Whether only synthetic or non-sensitive records are in scope |
For hosted inference, Vercel lists inclusionai/ling-3.0-flash-sante. For self-hosting, InclusionAI publishes the MIT-licensed inclusionAI/Ling-3.0-flash-Sante checkpoint. Verify the exact route your client can use. Check current provider terms before sending any data. See the DOM and vision for browser agents for a separate visual-page workflow.
Build a non-sensitive ticket handoff
Use a sandbox portal, synthetic fixture, or public demo account with no personal data. Ask the model to read the visible ticket list and return the ticket number, status, visible title, and page link. Do not ask it to infer urgency, diagnose a patient, decide treatment, or close a ticket.
| Output | Evidence | Review question |
|---|---|---|
| Ticket number | Exact visible identifier | Does it match the row? |
| Status | Exact visible status | Did the page finish loading? |
| Title | Short visible label only | Does it contain personal information? Remove it if so. |
| Source link | Ticket page URL | Can a reviewer open the same record? |
| Review flag | “Check manually” for ambiguity | Should this record be excluded? |
If the portal contains health information, stop before letting an external model or gateway process it. De-identification needs a real privacy and security review, not just a prompt instruction. Keep the first interaction read-only and follow your organization’s rules for access, retention, and human approval. The official Hugging Face model card warns against medical advice and requires qualified review for high-stakes use.
Prompt for a review-only ticket list
This is an illustrative workflow for synthetic or non-sensitive data. It is not a clinical-use procedure or an evaluation of model accuracy.
Use Tabbit CLI to read the support queue page I name in the signed-in test portal. The page contains synthetic or non-sensitive records only. Return each visible ticket ID, status, short title, and exact page URL. Do not infer medical urgency, provide medical advice, change a ticket, send a message, or close anything. Save the output as “Sante test queue — review”. Mark missing or ambiguous fields “Needs review” and stop before any write action.Expected output: one row per visible ticket, with the requested fields and a review flag. The reviewer should open each URL, verify the ID and status, and discard any row that contains sensitive content. If the client cannot select Sante or the provider is unavailable, do not silently substitute a similarly named model.
Use Tabbit Browser for the review step
Install Tabbit Browser on the same device as the local agent. Sign in to the test portal yourself. The agent can call Tabbit CLI in a separate task tab, leaving your normal browser tabs available for work. A fixture check in this worktree verified that a second CLI task could read a protected test page after a short-lived login cookie was set. It did not test Ling Sante, a Vercel request, or a clinical system.
Ask the agent to leave “Sante test queue — review” as a draft. Open it in Tabbit Browser, compare the ticket IDs and links with the source page, then decide whether to keep the draft. For general setup, see the Tabbit CLI guide, existing browser connection guide, first task checklist, and MCP versus CLI guide.
Limits and community evidence
Sante’s specialization does not make its output authoritative. The model card explicitly calls it a research model and lists hallucination, outdated information, and unsafe output among the limitations. A clinician or qualified reviewer must verify any high-stakes content against authoritative sources. This article stays with support-queue metadata and a human handoff.
The Reddit conversation reflects early interest in the available model lineup. It is a single community thread and says nothing about clinical reliability, local performance, or provider uptime. X search results did not expose readable posts in this browser session; YouTube search results were also unavailable. The draft therefore does not use social commentary as evidence for product facts.
Sources
For context beyond this task, see the Tabbit Browser overview, agentic browser explainer, and browser agent guide.
The same Reddit author also wrote, “I'd like more of these focused releases.” That is a preference for domain-focused releases, not a claim about browser performance or medical accuracy. Read the original discussion.
FAQ
Is Ling 3.0 Flash Sante a medical device?
No. InclusionAI labels it a research model, not a medical device. It must not be used on its own for diagnosis, treatment, or other high-stakes clinical decisions.
Can I use Ling 3.0 Flash Sante in a browser agent?
Only if your model host and local agent client both support it. Keep the browser task to synthetic or non-sensitive records and verify the provider terms before sending data.
Is Ling 3.0 Flash Sante closed source?
The worktree topic row labels it closed/API, but InclusionAI publishes downloadable weights with an MIT license. Hosted API availability and checkpoint access are separate routes.
Can it read real patient support tickets?
Do not send real health information unless your organization has approved the exact model host, data handling, and workflow. This guide uses only synthetic or non-sensitive ticket metadata.
Can Tabbit CLI reuse a website login?
A local Tabbit CLI task can connect to the Tabbit Browser environment where you signed in. A fixture test in this worktree verified session reuse, but it did not test Ling Sante or a healthcare portal.
What should the agent return?
Ask for visible ticket IDs, status, short titles, source links, and a review flag. A person should open each link and check the result before any write action.