MiniMax-M3 can help draft a source-linked brief from public search results and pages in a signed-in workspace, but the API does not provide a browser session. Use a local client that supports MiniMax-M3 and the Tabbit skill, then let Tabbit CLI reach the Tabbit Browser environment where you are already signed in. Keep public sources and private workspace sources labeled separately, and review the named draft before sharing.
Key takeaways
MiniMax AI lists
MiniMax-M3as an API model, and the publisher’s Hugging Face repository containsMiniMaxAI/MiniMax-M3weights with its own license. These are different access paths.MiniMax-M3.1-FlashandMiniMax-M2.7have separate IDs or repositories and terms; do not substitute them silently.API name
MiniMax-M3and repositoryMiniMaxAI/MiniMax-M3refer to different access paths.M3, M3.1 Flash and M2.7 are separate model routes and artifacts.
Check client skill discovery, CLI access, and the site login as separate steps.
Check the model route and access path
Use a local OpenCode session with a configured MiniMax provider; check the provider’s current model picker for its exact alias. The MiniMax API uses MiniMax-M3; a client-specific alias such as minimax-m3 is not the raw API ID. In OpenCode, check /connect and /models, then confirm Tabbit is discoverable as @tabbit from the project or personal skills. This is a client setup path, not a native MiniMax-to-Tabbit integration.
Separate the two source classes before summarizing
| Source class | What to capture | Review rule |
|---|---|---|
| Public web | URL, page title, visible date, exact supporting passage | Keep the original link; label estimates |
| Signed-in workspace | URL, title, visible status or date, relevant passage | Read only; do not copy unrelated private details |
| Comparison | Claim, source from each class, conflict | Mark unresolved differences for a person |
Draft a brief with source links
Open a small set of pages you are authorized to read. Check the Tabbit CLI guide, agent connection guide, and browser MCP versus CLI guide. Use public search for public pages and the local Tabbit CLI for the signed-in workspace. The CLI can reuse the existing login in the target Tabbit Browser environment; each task gets a separate tab so you can continue working. Parent-run local fixture evidence showed a protected page returning 401 before sign-in and 200 with heading “Release notes review” and record GEO-104 after a short-lived test cookie. This verifies CLI session reuse in that fixture, not a MiniMax integration.
Use the local Tabbit skill from this agent client. Do not submit forms or change source pages.
Create “M3 source brief, draft 01” from these public URLs and the specified signed-in workspace pages.
For every claim, include source class (public or workspace), URL, visible title/date, and a short supporting excerpt.
Separate confirmed facts, estimates, conflicts, and missing evidence. Do not merge conflicting claims.
Return a concise brief and a source table. Leave the draft for my review; do not share or publish it.| Checkpoint | Expected result | Stop and inspect when |
|---|---|---|
| Public-source pass | Search results and original URLs captured | A result lacks a primary source |
| Workspace pass | Authorized pages read in a separate task tab | Login is missing or scope is unclear |
| Synthesis | Each statement points to one or more sources | Dates or values conflict |
| Draft review | “M3 source brief, draft 01” open beside source pages in Tabbit Browser | Any claim cannot be checked |
Keep a visible review checkpoint in Tabbit Browser
Keep the draft open next to its sources in Tabbit Browser. Tabbit Browser keeps task tabs separate from your current browsing, so you can inspect source pages while the agent works. Review URL, title, visible date, and evidence for each claim. Correct or delete unsupported statements yourself. Do not give the agent permission to submit, publish, or alter the source workspace in this first pass.
MiniMax AI lists MiniMax-M3 as an API model, and the publisher’s Hugging Face repository contains MiniMaxAI/MiniMax-M3 weights with its own license. These are different access paths. MiniMax-M3.1-Flash and MiniMax-M2.7 have separate IDs or repositories and terms; do not substitute them silently.
For broader model information, read the existing MiniMax-M3 overview and the separate M3.1 Flash pricing page. This guide is specifically about browser evidence collection, source separation, and a human-reviewed draft.
More workflow guides: first task checklist, Tabbit CLI setup, Tabbit Browser explained.
References
MiniMax API overview — current API model names and access methods.
MiniMax M3 announcement — API availability and release context.
MiniMax M3 weights — repository and license metadata.
MiniMax M3.1 Flash weights and M2.7 weights — distinct artifacts and terms.
OpenCode models — provider-specific model selection.
Community notes
A Reddit user who tested eight prompts shared an individual comparison in the MiniMax M3 discussion. The observed feedback concerned coding tasks, not browser use; it is anecdotal and not a performance benchmark. See the broader M3 overview for model context.
See the first-task checklist and Tabbit Browser guide.
See the agent connection guide.
See the Codex Tabbit CLI setup.
FAQ
Can MiniMax-M3 browser use access my signed-in pages automatically?
No. The API model does not own a browser session. A local agent client must run Tabbit CLI against the already signed-in Tabbit Browser environment.
What is the exact MiniMax-M3 browser use API model name?
API: `MiniMax-M3`; weight repository: `MiniMaxAI/MiniMax-M3`. Check the license and current provider route.
Is M3 the same as M3.1 Flash or M2.7?
No. M3, M3.1 Flash and M2.7 have distinct model artifacts; confirm the exact provider route and license.
How should I handle conflicting source claims?
Treat each page as public or signed-in workspace evidence, preserve its URL and visible date, and label any disagreement.
What does the first browser task produce?
Create “M3 source brief, draft 01” with source links, then inspect the sources and draft in the browser before sharing.