Use Nemotron 3.5 Super VL with a signed-in browser only through a host that serves the exact vision-language checkpoint and a local agent that can run Tabbit CLI. The practical first task is to read a support ticket and its screenshot attachment, compare the ticket number and status, and stop before changing the record. NVIDIA released Super VL on October 15, 2026; it is a separate visual checkpoint from text-oriented Nemotron 3.5 Super (NVIDIA technical post). One community tester said their VL runs hit the default 16k-token output cap three times, so bound the task and check for truncation (test report).
Key takeaways
Use the VL model only when the page task needs image input, such as a screenshot or scanned attachment.
Verify the exact checkpoint and serving route. NVIDIA lists the model through NIM and publishes weights on Hugging Face.
Image input does not guarantee correct OCR, tool use, or access to a local browser.
Keep the first task read-only, preserve evidence URLs, and have a person compare extracted values with the visible source.
Do not conflate Nemotron 3.5 Super VL with the text-only Super release.
Check the model, host, and browser separately
| Layer | Check | Passing evidence | Failure path |
|---|---|---|---|
| Checkpoint | Select Nemotron 3.5 Super VL 120B-A12B, not text Super | Host shows exact ID and accepts an image input | Check model alias, modality support, and host version |
| Agent host | Can call the model and run local tools | One small inference and one skill discovery succeed | Separate provider/config failure from CLI failure |
| Browser | Tabbit task reaches the right signed-in ticket | URL, ticket ID, and account label are visible | Sign in yourself or stop at the login screen |
The research environment's pi --offline --list-models nemotron-3.5-super-vl-120b-a12b found no matching entry, and no inference was run. A listed model route is not proof your client can serve it. For the browser side, use the Tabbit CLI quick start, the CLI complete guide, and Codex connection steps. Pi invokes the Tabbit skill as /skill:tabbit; another client may use a different skill menu. A remote agent also needs a supported local handoff.
Extract the screenshot and ticket text as separate evidence
Sign in to the target support site in Tabbit Browser. In one owned task tab, ask the local agent to inspect the ticket text and attached image. Make it report each source separately, so a label read from the screenshot does not silently override the ticket record. A Reddit user reported that Nemotron 3.5 Super VL came closest in a niche game-rule test but still made a few factual errors; treat that as a single user's test, not a guarantee. It is a useful reminder to check the source for every value.
Try a bounded prompt:
Read the visible support ticket and its screenshot attachment in Tabbit Browser. Do not edit or submit anything.
Return the ticket number, status, and product label. For each value, say whether it came from page text or the image, include the current URL, and mark unclear text as unknown.
Name the review draft "Nemotron VL Ticket Review — GEO-219". Do not save it to the ticket.
Stop and show the extracted values beside the visible source so I can check them.GEO-219 is an illustrative test label, not an actual ticket or model run. The reviewer should compare the screenshot directly, confirm it belongs to the open ticket, and ensure a truncated response did not omit fields. If the image is unnecessary, use a text workflow instead of paying the cost of a visual checkpoint.
Keep the reviewer in control
| Field | Evidence to retain | Human checkpoint |
|---|---|---|
| Ticket number | Header text and attachment context | Confirm both refer to the same record |
| Status | Visible page text, plus image only if needed | Resolve disagreement in favor of the live record or stop |
| Product label | Screenshot crop and page URL | Verify spelling and context |
| Proposed action | Ticket request and approved policy link | Person decides whether to edit, reply, or close |
Use separate task tabs while you keep other pages open. Review how to connect an agent to an existing browser and the first-task checklist before a live ticket. Compare browser MCP and CLI or read the agent browser CLI tutorial if the host cannot launch the skill. Tabbit Browser is a practical local choice for a workflow that must use a login already open on your computer. Read the guides to browsers for agents, Playwright automation, and browser automation agents. The parent-run fixture proved only local session reuse, not Nemotron visual reading. Check the security risks of browser agents, then approve any change yourself.
FAQ
Is Nemotron 3.5 Super VL the same model as Nemotron 3.5 Super?
No. Super VL is the vision-language checkpoint intended to accept images; the text-focused Super checkpoint is separate. Verify the exact checkpoint name in your host.
Does Super VL guarantee accurate screenshot reading?
No. The model can misread text, crop context, or omit a field. Ask it to identify uncertainty and compare every important value with the page and image yourself.
Where can I run Nemotron 3.5 Super VL?
NVIDIA publishes the checkpoint and a NIM route. Local serving depends on a compatible inference stack, hardware, and configuration; a model card does not prove your machine can run it.
Can the model see my Tabbit login?
The model does not receive browser access by itself. A local agent calls a browser tool in the Tabbit environment where you signed in; check the page and account before each task.
Should an agent change a support ticket after reading its screenshot?
Keep the first run read-only. Have a person verify the ticket number, status, and requested action against the source before editing or sending anything.