Test Ornith 1.5 35B A3B on a signed-in browser task

To test Ornith 1.5 35B A3B on a signed-in browser task, keep the page and prompt fixed, change only the model endpoint, and use Tabbit CLI from a local agent. First verify that your runtime can serve the exact checkpoint and that the host can parse its tool calls. An endpoint that answers text prompts may still fail to open a browser page.

The official Ornith model card identifies a 36B-parameter checkpoint and documents Transformers, vLLM, SGLang, image input, and agentic usage. The A3B label is part of the model variant; it does not mean a 3B dense model. The page lists an MIT license. Serving notes do not prove that a chosen client supports the checkpoint’s chat template and tool-call parser.

Key takeaways

• Record the exact model ID, quantization, runtime version, and endpoint for each run.

• A working text request does not prove the host can parse browser tool calls.

• Install or discover the Tabbit skill, then test CLI launcher reachability separately.

• Sign in yourself in Tabbit Browser and begin with a read-only fixture.

• Compare returned record IDs and URLs with the visible page before deciding whether the model swap worked.

Check the runtime and tool-call path

Use an execution path documented in the model card, then check your own agent host. A tool-call parser can be the mismatch even when model serving is healthy. Avoid testing on a real ticket or account change; a harmless public page or synthetic queue is enough to confirm that the browser tool is actually invoked.

Set up a repeatable model-swap test

1. Download Tabbit Browser and open the profile for the authorized test site.

2. Sign in yourself and select a harmless read-only record. Do not copy credentials into the prompt.

3. Install or discover the Tabbit CLI skill. Pi invokes it as /skill:tabbit; other clients use their own skill list or picker.

4. Verify the installed local CLI is reachable from the same client environment. The local agent connection is different from skill discovery.

5. Run a baseline model, then Ornith. Keep the task and browser profile unchanged.

"$HOME/.local/bin/tabbit-cli" nodejs --task "Ornith fixture lookup" --request-id ornith-read-001 <<'JS'
await page.goto("https://example.com/", {waitUntil: "domcontentloaded"});
return await tabbit.observe({frames: "none", maxChars: 3000});
JS

Use a fresh, unique request ID for each new call. The command is the local launcher form; it is not a claim that your client already supports Ornith. For setup details, see connecting an agent to an existing browser, Codex Tabbit CLI setup, and browser MCP versus CLI.

Give Ornith one read-only task

Use the same named test record in both runs. Ask for visible fields and proof, not an interpretation. If the model text response looks correct but the browser tool trace is empty, treat it as a failed browser run.

Use the selected Ornith 1.5 35B A3B endpoint through this local agent and Tabbit CLI. In a separate Tabbit Browser task tab, open the signed-in test queue at [FIXTURE_URL]. Read ticket [FIXTURE_ID] only; do not reply or change it.

Return exactly: page heading, visible ticket ID, status, one linked source URL, and whether a browser tool call was made. If the endpoint cannot emit a tool call or the host cannot parse it, stop and report the runtime error.

Expected output: the five fields above, followed by “No write action taken.” This is a test fixture, not a production ticket.

Save the response as a named draft such as Ornith fixture lookup — run B. Open it in its separate Tabbit task tab and compare with the source record before copying anything. Your regular browser tabs remain available for other work. Follow the first-task checklist before allowing an agent to write.

Read the result as a compatibility report

A useful comparison reports whether each layer passed. If the text request succeeds but no tool call appears, inspect the host parser or template. If the tool call is emitted but not parsed, check the runtime version and agent framework. If the browser opens the wrong account or a login page, fix the Tabbit profile state before judging Ornith.

A community member who tested 21 Qwen3.8-27B quant variants on one RTX 5080 and a C-code workload said some variants were underwhelming. That is not evidence about Ornith or browser tasks, but it does show why comparisons should record exact weights and hardware. Read the original Reddit post.

See the agent browser tutorial, the browser research workflow, and web data to CSV for other bounded tasks. These explain the browser workflow; they do not imply Ornith is available in a particular provider.

What the local login check proves

A local Tabbit CLI check used a localhost fixture: it returned 401 before test sign-in, then a separate task tab read the heading “Release notes review” and fixture record GEO-104 after a short-lived test cookie was set. This verifies session reuse by CLI in that test environment only. No live Ornith endpoint, chat template, or host-side tool call was tested here.

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.