MiniMax describes MiniMax-M2.7-highspeed as a faster API version of M2.7 with the same results. It is a hosted speed route, not a second open-weight checkpoint you install beside the MiniMax-M2.7 weights. To use it with a signed-in browser, configure the exact highspeed ID in a local agent client that supports MiniMax and can run Tabbit CLI, then sign in to the target site in Tabbit Browser. The model endpoint, local tool access, and browser login are separate parts of the setup. MiniMax M2.7 model page
Community reports about M2.7 are mixed. One OpenCode user said M2.7 ignored rules injected through files and a plugin, while an OpenClaw user described it as useful for routine drafting and research but kept another model for harder reasoning. These are reports about M2.7 use, not a controlled highspeed browser test. OpenCode discussion · OpenClaw report
Key takeaways
MiniMax's official page lists two API IDs:
MiniMax-M2.7andMiniMax-M2.7-highspeed. It describes highspeed as faster with identical results.The highspeed route is hosted. MiniMax publishes one M2.7 weight repository; the model license is not an unrestricted commercial-use license.
No exact current third-party endpoint alias for highspeed was verified. Use MiniMax's exact ID through its own API route or confirm a provider's live listing first.
A local agent must both select that model route and invoke the local Tabbit CLI skill. Model availability alone does not provide browser control.
Compare routes on the same authorized test page, but report only observations from runs you actually made.
Which M2.7 route are you choosing?
| Route | Exact name to check | What the evidence supports | What it does not establish |
|---|---|---|---|
| MiniMax hosted standard API | MiniMax-M2.7 | MiniMax lists the standard API version | This is not proof of local weights or a specific agent-client integration |
| MiniMax hosted highspeed API | MiniMax-M2.7-highspeed | MiniMax says it returns identical results at faster speed; its token-plan price is higher than standard | Faster output does not prove more reliable page actions |
| Public weight repository | MiniMaxAI/MiniMax-M2.7 | Repository contains M2.7 code and model materials | It is not a separate highspeed checkpoint; model terms restrict commercial use |
| Normalized model catalog | minimax/MiniMax-M2.7-highspeed | A discovery/catalog string can help identify the model | It is not a verified API alias for every provider or client |
MiniMax's model page names the two API versions and says they have identical results, with highspeed running faster. Its token-plan pricing page lists a higher rate for highspeed than standard M2.7. MiniMax's Hugging Face model repository publishes M2.7 materials, but the repository's model license says commercial use requires prior written authorization.
Configure the model and browser separately
Start with a local agent client that has a documented MiniMax provider or accepts the official MiniMax API configuration. Check that the client's current model picker or provider settings accept the exact ID MiniMax-M2.7-highspeed. The normalized Models.dev key minimax/MiniMax-M2.7-highspeed is useful for catalog lookup, but it is not a promise that a provider accepts that string. I did not verify a current third-party highspeed endpoint or execute a model API request in this research.
Then set up the browser side. Install Tabbit Browser, sign in to the site in the intended profile, and check three separate conditions: the agent has discovered the Tabbit skill; its local process can run the installed tabbit-cli launcher; and the target site is signed in within the Tabbit profile. Pi documents /skill:tabbit. Other clients may expose /tabbit, /skills, or automatic skill loading. Confirm the syntax your client actually supports rather than assuming every client uses the same command. See the Tabbit CLI setup guide, existing-browser connection guide, and browser MCP versus CLI comparison.
A skill file does not prove that the client can invoke a local command, and a working CLI does not prove that a site is signed in. A remote coding agent or container also may not reach the desktop browser. Test the client and launcher on the same machine that runs Tabbit Browser. The first browser task checklist gives a read-only first run.
Compare both routes on the same page
The recipe below is a controlled comparison plan, not a measured M2.7 versus highspeed result. Use a test dashboard or a page you have permission to read. Keep the prompt, browser profile, page, baseline, and expected output fixed. Only change the model route between runs.
Use the currently selected MiniMax route with Tabbit CLI in a separate named task tab.
Open the authorized test dashboard page I provide. Do not change its data.
Compare the visible cards with baseline GEO-193 and report each changed field, its displayed value, timestamp, and source URL.
Record the exact model ID returned by the client, the page title and URL, the browser actions performed, elapsed time, and any errors.
Save a draft named "M2.7 highspeed dashboard rerun: GEO-193" and stop for my review.
Repeat the same prompt and checks after I change the route to the other MiniMax ID. Do not combine the two outputs.
Expected output for each run: route ID, page identity, action log, changes, unknowns, duration, errors, and draft link.GEO-193 and the draft name are illustrative. Do not infer a speed difference from one run or report a model ID unless the client shows it. A useful comparison is: same task and browser context, different explicit route, then a human check of both drafts. For general task boundaries, see best practices for Tabbit Browser and what an agentic browser does.
What the local browser check proves
In our local fixture check, the page returned HTTP 401 before a short-lived test cookie was set. A separate named Tabbit CLI task opened the same page in its own task tab and returned HTTP 200, the title Release notes review, and fixture record GEO-104. We removed the temporary cookie and scratch tabs afterward. This confirms local Tabbit CLI session reuse in that test environment. It does not show M2.7 or M2.7-highspeed making the browser calls, nor does it prove an external agent integration.
That separation is useful in practice. The model chooses or requests an action; the local agent runtime decides how to call tools; Tabbit CLI connects the task to the browser; the signed-in profile determines which pages are visible. The agentic browser overview and browser MCP versus CLI guide explain the difference between those layers.
Route and license checks before a real task
| If this is true | Do this next |
|---|---|
The client lists MiniMax-M2.7-highspeed from MiniMax | Select it and confirm the returned model ID before the browser task |
The client only lists minimax/MiniMax-M2.7-highspeed | Ask which provider receives the request and verify its current exact ID |
| The task uses downloaded M2.7 weights | Review the model license; do not assume commercial use is allowed |
| The agent can use the model but cannot run local commands | Hand the browser steps to a local agent or perform them yourself |
| The client, route, and signed-in page are all confirmed | Read the test page, save a named draft, and review it before any write |
Reddit users have reported both reliable and frustrating M2.7 behavior. One OpenClaw poster said it handled their routine drafting and research well, while another described failures to follow rules supplied through AGENTS.md and a plugin. A YouTube viewer also asked for a full setup guide under a video about M2.7 multi-agent use. These are useful signals about what people need from an agent setup, but none tests the highspeed route or Tabbit. OpenClaw discussion · OpenCode discussion · YouTube video and comments
Keep the first run read-only. If the highspeed route is unavailable in your chosen client, switch only after confirming the provider's exact standard route. Do not treat an ID from a catalog as proof of a working endpoint, and do not describe a comparison recipe as a benchmark. The goal is a sourced browser draft that you can check before changing the dashboard. For a related handoff that exports page data into a file, see the browser-to-CSV workflow.
FAQ
Is MiniMax M2.7-highspeed a separate open-weight checkpoint?
MiniMax presents M2.7-highspeed as a faster API version with the same results as M2.7. The public weight repository is named `MiniMax-M2.7`; the highspeed name is a hosted route, not a second set of weights documented by MiniMax.
What exact model ID should I configure?
MiniMax's model page lists `MiniMax-M2.7-highspeed` for the highspeed API and `MiniMax-M2.7` for standard M2.7. A catalog identifier such as `minimax/MiniMax-M2.7-highspeed` is not automatically a provider endpoint.
Does highspeed guarantee better browser actions?
No. MiniMax describes faster inference and identical results, not better browser-tool success. The agent client, tool schema, local runtime, page state, and permissions still affect the task.
Can M2.7-highspeed use my signed-in Tabbit browser?
It can be part of a workflow when a local agent client supports the exact MiniMax route and can run Tabbit CLI on the same machine as Tabbit Browser. The model API does not itself contain your browser session.
Is the M2.7 weight license unrestricted for commercial use?
The published model license permits non-commercial use under its terms and requires prior written authorization for commercial use.
How do I compare M2.7 and highspeed fairly?
Use the same page, prompt, browser profile, baseline, and expected fields for both routes. Record the returned model ID, actions observed, result, and duration; treat the comparison as unmeasured until you run it yourself.