To use Nemotron 3 Super 120B A12B with a signed-in webpage, first choose a model host that serves the exact checkpoint or API route, then connect a local agent client to Tabbit CLI. Ask the agent to read a small, named set of fields, place them in a draft table, and stop before submitting a form. The NVIDIA BF16 model card describes downloadable weights, not a guarantee that a particular API provider or agent client offers the model. Tabbit Browser can hold the signed-in page while a separate task tab keeps your normal browsing undisturbed. See the official Tabbit Browser download guide.
A LocalLLaMA user reported different results between vLLM and llama.cpp on a private benchmark, which is a reason to verify the serving path you actually choose (Reddit discussion). A separate hands-on YouTube first look includes a Browser OS test, but that demo does not establish access to a signed-in work account.
Key takeaways
Verify the exact model route with the host before building the workflow.
Do not treat 12B active parameters as the full BF16 memory requirement.
Keep model serving, client tool permission, and browser login as separate checks.
Ask for a named draft with source URLs, then review it yourself.
The local Tabbit fixture tested session reuse, not Nemotron inference or a model integration.
Check the setup at a glance
| Layer | Check | Stop if |
|---|---|---|
| Model host | Exact Nemotron route, access, and a short inference | Alias is missing or provider is unconfirmed |
| Agent client | Client can select the route and run local tools | Client is remote-only or local tool use is blocked |
| Tabbit Browser | Correct website is already signed in | The page shows a login challenge or wrong account |
Check the model and license before choosing a host
The official Hugging Face model card identifies the BF16 artifact as NVIDIA-Nemotron-3-Super-120B-A12B-BF16 and lists the NVIDIA Nemotron Open Model License. It shows Transformers loading instructions. That is a local-weight path, not the same thing as calling a hosted API. NVIDIA also lists the model in its NVIDIA NIM catalog, but a catalog entry does not prove that your account has a live endpoint. Check the host’s current model ID, region, quota, and terms.
The 120B total / 12B active naming is easy to misread. “Active” describes the parameters used for a token; it does not shrink the BF16 weight files to a 12B model. If you intend to run the weights locally, inspect the card’s files and hardware guidance first. If you intend to use an API, confirm the service and API identifier directly. The model’s open-weight license also does not mean every hosted service has the same terms.
The Tabbit CLI quick start explains the browser connection. In a compatible local client, install or discover the Tabbit skill using that client’s documented mechanism. Pi uses /skill:tabbit; some clients expose /skills or skill discovery rather than a named slash command. Skill discovery, CLI launcher reachability, and website login are three separate checks. Use the Tabbit CLI guide and existing-browser connection guide for those steps.
Read a signed-in page and prepare a draft
Open the target website in Tabbit Browser and sign in yourself. Do not send credentials to the model. Confirm the visible account and page title.
Start a separate local agent task and check that it can launch the installed Tabbit CLI. Use the client’s documented skill command.
Give the agent a narrow task. For example, extract the visible item name, date, and status from one page and retain a source URL for every value.
Ask for a draft table named
Nemotron intake review — GEO-202. Keep it in the agent’s response or a separate draft document. Do not let the agent write into the source form.Compare each row with the page. Mark missing or ambiguous values
unknown; do not fill them by inference.Enter any approved values yourself and submit only after a human review.
A useful prompt is:’
In the signed-in page open in Tabbit Browser, read only the visible record fields Name, Date, and Status. Do not click buttons, edit, or submit.
Return a table with columns: field, exact visible value, source page URL, confidence. Use “unknown” for missing or ambiguous values.
Prepare a separate draft named “Nemotron intake review — GEO-202”. Stop before writing to the website. Show me the source page and draft so I can compare them.| Page condition | Agent output | Human check |
|---|---|---|
| Text is visible in the page | Exact text with its page URL | Confirm it belongs to the intended record |
| Field is absent or unclear | unknown | Find the field or leave it blank |
| Value appears only in an image | No inferred value | Inspect it manually or use a separately verified vision model |
| Form has a submit control | Draft values only | Check account, values, and destination before submitting |
Put the browser step in a separate Tabbit task
Tabbit Browser is useful here because the signed-in page stays in the browser while the agent works in a separate task tab. You can keep another tab for normal browsing. The CLI can reuse the browser session available to that local environment, but this depends on using the intended Tabbit instance and profile. A remote agent runner cannot reach your local browser unless it has a separately supported local handoff.
A local test fixture checked this boundary: a protected page returned 401 before sign-in, then a second task using the same browser session reached the fixture after a short-lived test cookie was set. The page showed “Release notes review” and fixture record GEO-104. This verifies session reuse by Tabbit CLI in that fixture. It does not verify Nemotron inference, a named provider, or automatic compatibility with any agent framework.
Use the first-task checklist to keep the first run narrow. If your client cannot invoke the local CLI, compare browser MCP and CLI or see the browser agent CLI tutorial. Review browser-agent security guidance before granting access to sensitive pages. For more context, see the browser automation overview.
When this workflow is a good fit
Use it when the exact model route is available in a local-capable client, the page is already signed in within Tabbit Browser, and the task can stop at a reviewable draft. It is a poor fit if you need an unattended transaction, the agent must read image-only content, or the selected host and model route remain unverified. Start with one low-risk page and compare every extracted value before you expand the task.
FAQ
Can I use the Hugging Face BF16 weights as an API model?
No. The Hugging Face card describes downloadable model weights. An API route is a separately hosted service. Confirm the exact model and access with the provider you intend to use.
Does 120B mean the model needs only 12B parameters in memory?
No. The card names a 120-billion-parameter model with 12 billion active parameters. Active parameters per token do not make the full BF16 checkpoint a 12B download.
Can Nemotron 3 Super control a browser by itself?
A model does not connect itself to a browser. The agent client must support the model and permit a local browser tool such as Tabbit CLI.
Can it submit a form after reading the page?
It can prepare values for review, but this workflow stops before submission. A person should check every field and decide whether to submit.
Can I use the model in Pi?
Pi can invoke the Tabbit skill with `/skill:tabbit`. This research did not verify a provider that serves this exact model through Pi, so check your own model endpoint first.