Nemotron Cascade 2 can help plan a browser task, but it does not log into websites or connect to Tabbit on its own. For a signed-in page, use a local agent client that can run the model and local tools, connect that client to Tabbit CLI, sign into the site in Tabbit Browser, and ask for a source-linked draft that you review before sharing. NVIDIA publishes open weights and local inference instructions; its Hugging Face page currently says the model is not deployed by an inference provider. NVIDIA model card · Tabbit CLI quick start
Community reports are mixed. One commenter said it was "very, very fast" for coding; another reported loops and timeouts in agentic tasks. Treat that as a reason to keep a first browser run small and inspect every step, not a verdict about all runtimes. LocalLLaMA discussion
Key takeaways
Treat Nemotron Cascade 2 as the reasoning model. The agent client supplies tools, and Tabbit CLI connects browser actions to Tabbit Browser.
NVIDIA identifies the model as open-weight, released March 19, 2026, under the NVIDIA Open Model License. Open weights do not mean an unrestricted license or a ready-made browser integration.
The model card shows local Transformers, vLLM, and SGLang routes. It currently reports no Hugging Face Inference Provider deployment.
Check exact model access, tool-call parsing, and local launcher availability separately from the website login.
Start read-only. Ask for a named draft, citations next to claims, and a human review checkpoint.
Choose a route before opening the workspace
| Need | Route to evaluate | Check first |
|---|---|---|
| Run released weights | Local inference with vLLM, SGLang, or Transformers | Hardware, runtime/model revision, chat template, quantization, and tool-call parser |
| Use a hosted endpoint | Provider listing the exact model ID | A live endpoint; HF currently says no Inference Provider deployment |
| Use browser tools | Local agent client plus Tabbit CLI | Client can run local tools and parse calls from the configured model |
| Keep a site session | Tabbit Browser | Correct profile and site permissions; model access grants no account access |
| Use a remote-only runner | Supported local handoff, if available | Otherwise give the source checklist to a local operator |
The released model has roughly 32B parameters even though A3B refers to 3B active parameters. Do not estimate hardware needs from active parameters alone. NVIDIA's model card links runtime and tool-use instructions; check your quantization and context length before downloading or allocating compute. Model card and runtime examples
The card documents a custom XML-style tool-call format and an agent example. That shows how the model can express a call; it does not prove every client parses it. For a browser task, the client must expose browser tools and translate the model output correctly. NVIDIA tool-use example
Connect the model, agent, and browser separately
Install and open Tabbit Browser, sign into the target site yourself, then start a local agent client. The official Tabbit CLI guide documents the Tabbit skill for supported clients. Tabbit's own quick start uses /tabbit for its tested clients. Pi uses /skill:tabbit, and some clients expose /skills or load skills automatically. Follow your client's current docs instead of assuming a slash command works everywhere.
Check three things independently:
Skill discovery: the agent can find the installed Tabbit skill and its client-specific invocation.
Launcher reachability: the local
tabbit-clicommand runs from the agent environment.Site authorization: the signed-in Tabbit Browser profile can access the intended page.
A model API response does not prove local process access, tool-call parsing, or browser account access. See connecting an agent to an existing browser and browser MCP versus CLI for connector choices.
Make a small source-linked brief
Use a separate task tab for the agent so you can keep working in the browser. Ask it to visit only named public pages and the specific signed-in workspace page needed. Label each source, record what it supports, and leave the output as a named draft. The first-task checklist is a read-only starting point.
Use the installed Tabbit skill from this local agent. In a separate task tab, open only the public sources and signed-in workspace pages I name. For every page, record PUBLIC or WORKSPACE, URL, visible title, date/version, and one claim it supports. Do not follow unrelated links, edit source pages, or copy credentials. Mark inaccessible pages and unsupported details as unknown.
Create a draft named "Cascade browser brief — GEO-190" in the selected workspace. Include a short summary, one source-table row per page, and open questions. Put each URL beside the claim it supports. Do not publish or share it. Stop and show me the draft.
Expected output: draft link, selected workspace, separate PUBLIC and WORKSPACE rows, cited claims, and unknowns.GEO-190 is an illustrative draft name. Check the workspace and citations yourself. Do not let the model submit forms or change records in its first run. If it loops or returns malformed output, stop, save visible evidence, and retry with fewer pages or a runtime and parser you have verified.
Read the result before widening the task
| Observation | Next step |
|---|---|
| Exact endpoint answers and the client parses a test call | Try one read-only public page, then one authorized workspace page |
| A hosted catalog lists the name but no endpoint is documented | Treat availability as unverified; use documented local routes or another verified model |
| Output repeats calls, loops, or ignores a stop condition | Stop; reduce tools and pages, inspect runtime and parser settings |
| Signed-in page asks you to log in or denies access | Sign in yourself in the intended profile or request access; do not bypass it |
| A claim has no supporting URL | Remove it or mark it unknown before sharing |
| Agent cannot reach local tools | Use a client with a supported local connection or hand off to a local operator |
The Tabbit CLI complete guide, agent connection guide, first browser task checklist, and Codex setup guide cover setup choices. For broader decisions, read browser automation agents, agent browser setup, Codex browser versus Tabbit CLI, and browser agent security risks. Public search cannot establish private workspace state; web search versus browser control explains the difference.
FAQ
Can Nemotron Cascade 2 open my signed-in browser by itself?
No. The model produces text or structured tool calls, while a separate agent client and local browser connector provide tools. Tabbit CLI can connect a supported local agent to Tabbit Browser after you install its skill, confirm the CLI launcher, and sign into the target site in that browser.
Can I run Nemotron Cascade 2 locally?
NVIDIA publishes the weights on Hugging Face under the NVIDIA Open Model License. The model card documents Transformers, vLLM, SGLang, and compatible local applications, but hardware, runtime version, quantization, and context size determine whether a setup is practical.
Is Nemotron Cascade 2 available as a hosted model?
The Hugging Face model page currently says it is not deployed by an Inference Provider. Check the exact model ID and endpoint again before choosing a hosted route; a catalog entry alone does not prove that an API is callable.
Does the model's tool calling automatically control Tabbit?
No. NVIDIA documents a tool-call format and examples, but a client must translate the model output into tools it can run. The model, client adapter, Tabbit CLI, and browser session are separate parts that need to work together.
How do I keep browser research safe to review?
Start with read-only pages and ask the agent to record each URL, title, date, and supported claim. Have it create a named draft with unknowns and citations, then check the draft yourself before editing or sharing it.
Can I keep working while a Tabbit task runs?
Yes. Tabbit's CLI guide says a task uses the existing signed-in browser state while you can continue in other windows. Keep the agent in its separate task tab and return to inspect its draft and source links.