Mistral Medium 3.5 with a signed-in browser: check the route first

Mistral Medium 3.5 is a 128B open-weight model released as version 26.04, but Mistral’s model page now marks it deprecated with an October 31, 2026 deprecation date and recommends Large 4 for new integrations. The legal model registry still labels it Active, so its current serving status is unclear. If your account or local runtime still exposes the exact fixed ID mistral-medium-3-5, you can use a local Mistral Vibe CLI agent with Tabbit Browser to compare public sources with an already-authorized team page and save a reviewable brief. Verify the route before starting.

This is a browser workflow, not a model review. It fits a short research task where public product information must be separated from a private workspace note. The open weights are published under a Modified MIT license; that does not make the hosted API, model availability, or Vibe account access interchangeable. For local browser work, use Vibe CLI on the same machine as Tabbit Browser. Mistral’s remote Vibe Code Web runs in a managed cloud sandbox and needs a separately supported local handoff to reach your browser.

Key takeaways

• Confirm the live fixed ID; the dated deprecation record conflicts with the legal Active listing.

• Do not replace mistral-medium-3-5 with the rolling latest alias.

• Open weights, hosted access, Vibe entitlement, and local hardware are separate checks.

• Use local Vibe CLI with Tabbit Browser and stop at a reviewable draft.

Resolve the version and availability question first

Mistral’s model catalog gives the dated version as 26.04 and the stable fixed API ID as mistral-medium-3-5. Its lifecycle guide explains that a major-minor ID is fixed, while mistral-medium-latest can move to a later Medium release. The current model card says the model is deprecated, gives October 31, 2026 as the date, and directs new integrations to Mistral Large 4. The Mistral Legal Center registry still labels Medium 3.5 Active. Those official records conflict on present status. Test the exact ID with your own account or self-hosted runtime before designing the task around it.

Sources: official model card, model lifecycle policy, Legal Center register, and Vibe CLI skills.

Keep public research separate from the signed-in workspace

A Reddit user described Medium 3.5 as strong for coding and agentic tasks based on their own trial. Another LocalLLaMA user said a few hours of testing found issues their usual model missed, while discussing a 3-bit local quantization. These are personal reports from different setups, not evidence for browser accuracy or a reason to treat the weights as easy to run. MistralAI discussion · LocalLLaMA discussion.

Use this prompt after confirming the route. Replace bracketed values with authorized inputs:

Use Mistral Medium 3.5 only if the exact fixed model ID `mistral-medium-3-5` appears in the provider’s live model list for this account or local runtime. Do not substitute `mistral-medium-latest`. From the local Mistral Vibe CLI, invoke the installed Tabbit skill as `/tabbit` if the skill is discovered by this host. Confirm Vibe CLI can run local tools, then use Tabbit CLI to create a separate task tab in the existing Tabbit Browser session. Read only [public source URLs] and this authorized workspace URL: [workspace URL]. Compare [stable record ID] and [field]. Keep public facts and internal notes in separate sections with exact source URLs, visible dates, and quotations short enough to verify. Create a browser-visible document titled “Mistral 3.5 source brief — draft 01” with a “Needs human review” heading. Save as a draft, reopen it, and verify the title, draft status, and source links. Do not send, publish, or change source records. If the model route is unavailable, stop and return the source checklist without changing aliases.

Use Vibe CLI locally when the browser must stay on your machine

Mistral documents Vibe Code as a local CLI, VS Code extension, or remote web session. The CLI is the right host to evaluate for this recipe because it runs in a local shell and loads skills; Mistral documents a user-invocable skill as /<skill-name>. Confirm the tabbit skill appears in Vibe before using /tabbit. Then separately confirm the model route, local tool permission, Tabbit CLI connection, and the website login. The hosted Le Chat / Vibe experience is not the same thing as the local CLI. The model may request an advertised function, but the local host executes it; this is not native Tabbit browser access. A cloud runner cannot reach a local browser without a supported handoff.

The assigned local fixture showed a protected page returning 401 before sign-in, then 200 from a second named Tabbit CLI task after a temporary test cookie was set. Each task owned one task tab. This verifies local Tabbit CLI session reuse in this environment. It does not verify Mistral Vibe, Medium 3.5, API availability, or a model-to-Tabbit connection. The useful result is the browser-visible draft reopened for review with its sources intact.

Related local steps: Tabbit CLI setup, connect an agent to a signed-in browser, first-task checklist, browser MCP vs CLI, local versus cloud agents, research with source links, safe login reuse and Codex local setup example.

Open weights do not settle the browser setup

Mistral calls Medium 3.5 open-weight and publishes it under a Modified MIT license. That tells you weights are available subject to the license; it does not tell you that the hosted API still accepts requests, that a Vibe plan includes the model, or that local weights fit your hardware. Choose hosted API only after a live account check. Choose self-hosting only after checking the published license, runtime support, and hardware. If Medium 3.5 is retired for your route, use a currently supported model and keep the same Tabbit review workflow; do not silently let a rolling alias change the task.

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.