To use Ornith 1.0 35B with a signed-in dashboard, run it through a compatible local agent, connect that agent to Tabbit Browser with Tabbit CLI, and review a read-only change report before making edits. The model card documents tool calling and OpenAI-compatible serving, but that does not prove a particular agent client works with your local endpoint. Check the model runtime, agent connection, CLI launcher, and site login as separate steps. Official model card · Tabbit CLI quick start
A Reddit user who tested 35B on coding tasks reported that it sometimes changed code beyond the requested edit, which made review cycles necessary. That is one person's coding experience, not a browser benchmark. It does support a practical rule for any page workflow: ask for observed facts and a draft first, then verify each proposed change yourself. Original Reddit post
Check the 35B serving path before connecting a browser
The release describes Ornith 1.0 35B as MoE and 31B as Dense, so these are distinct variants, not aliases. However, the current official Ornith-1 checkpoint table lists 9B, 35B, and 397B; it does not list 31B or provide a 31B card. Treat the announced 31B variant as unavailable until an official downloadable checkpoint and serving instructions can be confirmed. For 35B, the model card lists recent runtime versions and an eight-GPU, 80 GB per GPU serving recipe. It also describes the model as suitable for single-GPU deployment elsewhere, so its hardware notes conflict. Check the exact quantization and runtime before choosing hardware. The card lists MIT licensing and OpenAI-compatible tool calls. Ornith-1 official repository · Ornith 1.0 35B model card
| Layer | Check | What a pass means |
|---|---|---|
| Checkpoint | Use the exact ornith-ai/Ornith-1.0-35B or equivalent published 35B checkpoint and confirm its license | You know which variant and files you are serving |
| Runtime | Match the model card's Transformers, vLLM, or SGLang requirements; enable the documented tool-call parser | The endpoint can return structured tool calls |
| Agent client | Confirm the client can call your local OpenAI-compatible endpoint and expose local skills | A harmless test tool call returns structured arguments |
| Browser | Open the target dashboard in Tabbit Browser and sign in yourself | The intended account and page are visible |
| Review | Ask for a read-only change list with links or row labels | Each claim can be checked against the page before action |
At present, the model card verifies model and serving routes. It does not document Ornith plus Tabbit CLI as a preconfigured integration. Check your client’s provider support before you choose a hosting service or model endpoint. If the agent runs remotely and cannot reach the local launcher, have it prepare a task list for a local agent instead.
Connect the local agent to the signed-in Tabbit environment
Install Tabbit Browser and open the dashboard you are allowed to inspect. Sign in yourself. In the local agent, discover the installed Tabbit skill through the client’s actual skill interface. Pi documents /skill:tabbit; some clients expose /skills, while others load installed skills without a named command. Then check that the tabbit-cli launcher runs on this machine. Skill discovery confirms the instructions are available. Launcher reachability confirms local control works. Opening the site in the target Tabbit profile confirms the website session.
Tabbit CLI can create a separate task tab in the current browser environment, so you can keep working in your own tab. The CLI can reuse the signed-in session where the site permits it; it does not sign in for you or grant additional account permissions. See the existing browser connection guide and first browser task checklist.
Ask Ornith to report dashboard changes, not make them
Choose a dashboard with a visible update log or a test workspace. Ask the agent to read a bounded date range, summarize changed rows, and mark uncertain fields. Do not give it permission to save, submit, invite users, or change access while you are still assessing its output.
Use the Tabbit skill installed for this local agent. Open the dashboard I already signed in to, in a separate task tab. Read only the entries dated 2026-11-09. Do not click edit controls, save, submit, invite anyone, or change access.
Prepare a draft named "Ornith 35B dashboard delta review — ORN-35B-017". For each reported change, include the visible row label, previous value if shown, current value, date, and the page URL. If a field is not visible, write "unknown". Keep possible corrections in a separate section named "Needs human check". Do not apply or publish the draft.
Expected output: a concise change table, source row labels, unknowns, and a draft link for my review.ORN-35B-017 and the draft are illustrative. The task uses no credentials in the prompt. Keep the named draft open for review, compare each row label and value with your own tab, and ask the agent to explain any mismatch before approving an edit.
Review the output against the page
| Agent result | Your check | Next step |
|---|---|---|
| Every row label and value is visible | Compare the cited row and date in your own tab | Accept the report as a draft; decide separately whether to edit |
| A value is absent or ambiguous | Find the row in the page and inspect its context | Keep it under “Needs human check” or mark it unknown |
| The page redirects to sign-in or shows another account | Check the selected Tabbit profile and target account | Stop; sign in yourself or select the intended profile, then retry read-only |
| Tool arguments are malformed or the agent proposes a write | Check the model parser and client tool mapping | Do not retry a write; inspect configuration with a harmless call |
| The agent cannot reach the local CLI | Check the local launcher and client runtime | Have a local agent perform the browser step; use a remote agent only for a handoff plan |
One X model-evaluation write-up described treating the model’s first output as something to check against a test harness before using it. That report tested code-agent behavior, not browser interaction, but the same review habit fits this dashboard task. X report
For setup details, read the Tabbit CLI complete guide, MCP versus CLI comparison, and Codex Tabbit CLI setup. The browser agent setup guide, browser automation overview, and browser agent security guide cover client selection and access boundaries. The web search versus browser control guide explains why a search result cannot confirm a change inside a logged-in page. None of these links claims that this specific Ornith deployment was tested with Tabbit.
FAQ
Can Ornith 1.0 35B open my signed-in browser by itself?
No. Ornith is the model, while a local agent and browser tool perform the page actions. Connect a compatible agent to Tabbit CLI and sign in to the target site yourself.
Is Ornith 1.0 35B the same checkpoint as Ornith 1.0 31B?
They are distinct announced variants: 35B MoE and 31B Dense. The current official Ornith-1 checkpoint table lists 9B, 35B, and 397B, but no downloadable 31B checkpoint or model card. Do not assume a runnable 31B route or infer its hardware needs.
What runtime does Ornith 1.0 35B need?
Its model card lists Transformers 5.8.1 or later, vLLM 0.19.1 or later, and SGLang 0.5.9 or later. The documented server recipe uses an eight GPU, 80 GB per GPU node, so check your exact deployment plan before downloading.
Does using Tabbit CLI give Ornith my website password?
No. The model does not receive the browser's password or cookies. A local Tabbit task can work in the signed-in browser environment, subject to the site account's existing permissions.
How do I find the Tabbit skill in my agent client?
Use the method documented by your client. Pi uses `/skill:tabbit`; other clients may use `/skills` or load installed skills automatically. Then confirm that the local Tabbit CLI launcher is reachable.
What should I verify before accepting a dashboard summary?
Check that each reported change points to a visible row or field, and have the agent list unknowns separately. Treat any suggested update as a draft until you compare it with the page and approve it yourself.