As of November 10, 2026, Ornith 1.0 31B is an announced Dense variant, but I could not verify downloadable weights or current serving instructions in the official Ornith-1 checkpoint list. Do not attempt a model swap using a guessed ID. You can still prepare a controlled comparison: run a read-only baseline in Tabbit Browser with a model endpoint you have already verified, then mark the 31B run as blocked until its official checkpoint and runtime are available. Ornith release repository · Release announcement
That caveat matters because the family announcement distinguishes 31B Dense from 35B MoE, while the current official checkpoint table lists 9B, 35B, and 397B for version 1.0. A model catalog listing or community quant is not the same as a verified official artifact. Official checkpoint list
Verify what is actually available before the swap
Keep model availability separate from browser access. First check for an official 31B checkpoint page, downloadable files, license, and serving instructions. Then check whether your client can call that model through a supported local runtime. Only after both pass should you run the second comparison. The family announcement says 31B Dense and 35B MoE are distinct variants; it does not establish a 31B host or hardware requirement. Ornith release note
| Check | Evidence to find | Current decision |
|---|---|---|
| 31B identity | Official release describing the 31B Dense variant | Announced separately from 35B MoE |
| 31B artifact | Official model card and downloadable checkpoint files | Not listed in the current official 1.0 checkpoint table |
| Runtime route | Official serving instructions and a client that supports the endpoint | Unverified for 31B; do not guess a provider or model ID |
| Browser route | Tabbit Browser signed in to the authorized dashboard; local Tabbit CLI reachable | Test independently with an already verified baseline model |
| Comparison permission | Read-only task with no save, submit, or access changes | Human reviews the report; no write action during the test |
The Tabbit CLI quick start explains the local browser connection. It does not list Ornith 1.0 31B as a tested integration. Keep that distinction visible in your notes.
Record a baseline on the same signed-in page
Install Tabbit Browser and sign in to a dashboard you are authorized to inspect. In your local agent, use the client’s actual skill entry. Pi documents /skill:tabbit; some clients expose /skills, and others load installed skills automatically. Check that the tabbit-cli launcher works separately. Then open the same allowed page in a separate task tab. You can continue working in your own browser tab while the task runs.
Use a model endpoint that has already passed your local tool-call check. This could be a verified 35B deployment if its runtime and hardware are already working for you, or another model endpoint your agent client supports. The 35B card lists vLLM and SGLang routes, but its hardware language is inconsistent; do not use it as a casual setup recommendation. If the local endpoint is unavailable to the agent client, stop before the browser step. A remote runner that cannot reach the local CLI can prepare a handoff checklist, but cannot perform this local baseline.
For each run, preserve the same page, date range, prompt, permissions, and output fields. Record the model ID exactly as returned by your verified endpoint, artifact or quantization, runtime version, agent client, and timestamp. Do not compare a quantized 35B baseline with an imagined 31B result.
Use one prompt, and keep a missing run visibly blocked
The prompt below creates a read-only baseline record. It explicitly prevents the agent from filling in missing 31B results. ORN-31B-021 is an illustrative draft name, not a real test record.
Use the Tabbit skill installed for this local agent. In a separate task tab, open the dashboard I have already signed in to. Read only the entries dated 2026-11-09. Do not edit, save, submit, invite users, or change access.
Prepare a draft named "Ornith 31B model swap baseline — ORN-31B-021". Record the baseline model identifier exactly as this agent reports it, artifact or quantization if known, runtime, agent client, date, page URL, visible row labels, and observed values. Do not guess missing fields.
The announced Ornith 1.0 31B Dense variant is not in the official checkpoint table I verified. Add a second comparison row with status "BLOCKED: official downloadable checkpoint and runtime not verified". Do not call a 35B model 31B, invent a 31B model ID, or fill in comparison results. Leave the draft unpublished and show it to me.
Expected output: one measured baseline row, one explicitly blocked 31B row, unknown fields, and a draft link for human review.Tabbit CLI can create a separate task tab in the existing Tabbit environment and reuse a signed-in session where the site allows it. This confirms browser access, not model availability. The existing browser connection guide and first browser task checklist explain the local session checks.
Compare results only after the second checkpoint is verified
| Comparison field | Keep constant or record | Review rule |
|---|---|---|
| Page and account | Same URL, authorized account, and visible date range | Stop if the page or account changes |
| Prompt and permissions | Same prompt and read-only scope | Do not allow one run to make edits |
| Model setup | Exact checkpoint, quantization, runtime, endpoint, and client | Mark unknown details instead of inferring them |
| Page observations | Row labels, values, omissions, and uncertainty | Compare with what is actually visible |
| Result status | Baseline measured; 31B pending or blocked until its official route is confirmed | Never present the blocked row as a model result |
A Reddit user discussing open models noted in June that the 31B Dense weights did not appear to have been released. That is an early community observation, not the current source of truth; the official repository check above remains the deciding evidence. Another local-model user recommended trying a specific quantization for an agentic coding workflow, a reminder that the artifact and quantization need to be recorded when comparing runs. June discussion · Local agentic-model discussion
For connection details, see 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 local access and permissions. These guides do not establish that Ornith 1.0 31B weights or a provider are currently available.
FAQ
Can I download and run Ornith 1.0 31B today?
The release announcement identifies a 31B Dense variant, but the current official Ornith-1 checkpoint table does not list its weights or a model card. Confirm an official download and runtime guide before treating it as runnable.
Is Ornith 1.0 31B another name for Ornith 1.0 35B?
No. The announcement describes 31B Dense and 35B MoE as separate variants. A model catalog entry is not evidence that both have downloadable weights or an available provider.
Can I still test a browser workflow while 31B is unverified?
Yes. Use a model endpoint you have already verified for a read-only baseline in Tabbit Browser. Record that result, mark the 31B run as blocked, and do not invent a second result.
Does the local baseline share my website login with Ornith?
The local Tabbit task can use the signed-in browser environment when the site permits it. The model does not receive your password or cookies, and the task does not add account permissions.
How do I invoke the Tabbit skill for the comparison?
Use the client’s documented skill method. Pi uses `/skill:tabbit`; other clients may use `/skills` or load installed skills automatically. Also verify that the local Tabbit CLI launcher is reachable.
What makes two model runs comparable?
Use the same page, date range, prompt, permissions, and output fields. Record the exact model artifact, quantization, runtime, and agent setup for each run, then have a person verify the page observations.