To prepare form fields from a screenshot with MiniCPM-V 4.7 35B-A3B, install Tabbit Browser and use Tabbit CLI from a local agent that can reach a supported vision runtime. Let the model read the fields. Keep the proposed values in a named draft you can review beside the source page. OpenBMB model card
A safe field handoff
Open the target page in Tabbit Browser and confirm you are signed into the intended account and environment. In a local agent that can run tools on this machine, install the Tabbit skill using the client’s documented mechanism. Use /tabbit only in clients that expose that named skill; Pi uses /skill:tabbit, while some clients expose /skills. First confirm the client can select or call the exact MiniCPM-V 4.7 35B-A3B runtime. OpenBMB publishes weights and inference examples, but that does not mean every agent client can load the vision model or invoke local tools.
The agent should read the source page, map each field to evidence, fill a separate draft, and stop before Save or Submit. A practical checkpoint is a named draft such as GEO-104-review, with a visible table of extracted values and source labels. Keep the CLI task in its own tab context; you can continue browsing elsewhere. A remote runner cannot reach this local browser unless it offers a separately verified local handoff.
| Decision | What to check | Stop condition |
|---|---|---|
| Model route | Exact 35B-A3B variant appears in the local runtime | No confirmed model ID |
| Page evidence | Each value is visible in the signed-in page | Hidden or ambiguous field |
| Form state | Draft record is separate from live submission | Submit button or irreversible action |
| Input | Expected output | Human check |
|---|---|---|
| Visible invoice or ticket page | Field, value, source text, confidence | Match each value to the page |
| Missing or blurred value | unknown | Do not infer it |
| Destination form | Unsaved draft only | Confirm account and environment |
A useful way to work with the result is to ask the model to revise a named browser draft, then inspect the visible evidence before approving the page change. For the actual connection steps, see the Tabbit CLI tutorial, existing-browser connection guide, setup checklist, MCP and CLI comparison, Codex setup, Claude Code setup, the Haiku guide, and the Sonnet guide.
MiniCPM-V 4.7 is published in 35B-A3B and 1B variants. Hugging Face lists the 35B-A3B card as image-text-to-text under Apache-2.0 and provides Transformers and vLLM routes. These facts describe model files and supported inference paths, not an agent integration or browser automation guarantee. The model card does not establish that the model can call Tabbit CLI, navigate a live website, or reliably map page values to form fields.
A Reddit discussion about the mobile demo says the 35B-A3B variant caught the poster’s attention for reading signs and documents. Another LocalLLM post calls the reported low-memory setup interesting but flags the large download and device-memory requirements. These are user reactions and setup discussion, not accuracy evidence. I found no relevant X post; a YouTube result was unavailable.
If your runtime cannot prove the exact variant and image input path, do not start with a real form. Use a local screenshot fixture and a non-production test page, then compare each proposed value with the visible source. Keep sensitive page content inside your approved environment. The model’s vision capability does not grant access to the site, and site login does not prove the model route works.
Use MiniCPM-V 4.7 35B-A3B only if this local agent runtime confirms that exact model is available.
Read the currently open, signed-in page. Extract only these fields: [field names].
For each field return: value, exact visible source text, and status (verified / ambiguous / unknown).
Do not use memory or infer missing values. If a value is unclear, return unknown.
Fill a separate draft form with the verified values. Do not click Save, Submit, Send, or any control that changes the live record.
Name the draft GEO-104-review. Show the draft and the evidence table, then wait for my review.This is an editorial example, not a tested MiniCPM or agent prompt.
Where Tabbit CLI fits
Tabbit Browser is useful when the task depends on an already authenticated webpage and a person needs to keep working in the browser. Tabbit’s public CLI materials describe browser tasks from supported local agents and reuse of the browser login state. The supported path must still be installed for the chosen client. The current project research verifies this behavior with a local test fixture; it does not verify MiniCPM-V integration or third-party model hosting.
Use this pattern for low-risk data preparation where a person checks every field. Do not use it to approve payments, send messages, modify production records, or decide ambiguous legal, financial, or medical information. Before a real action, confirm the account, destination, field evidence, and live/ test environment in the browser.
FAQ
Can MiniCPM-V 4.7 35B-A3B control a browser on its own?
No. It is a vision-language model. A separate local agent runtime must expose the model and browser tools.
Can I run it in a local agent?
Only if that client supports the exact model variant, image inputs, and local tool execution. Check the model picker and runtime documentation first.
Does Tabbit CLI inherit the site login?
A local agent using the CLI can work with the login in its target Tabbit browser environment. Confirm the correct profile and website before each task.
Can it submit the form?
This workflow deliberately stops at an unsaved draft. A person should inspect the evidence and make any consequential submission.
What should I do if the source value is unclear?
Leave the proposed field blank and show the source for manual review.
Has this prompt been tested with MiniCPM-V?
It is an example. Use an authorized test page to verify the exact model, runtime, and agent.