LightOnOCR-3 0.8B cannot operate a signed-in browser. It is an OCR and document-reading model. Use it on a locally saved screenshot or PDF page, then hand its extracted text to a separate local agent that can run Tabbit CLI. That second agent can inspect the matching signed-in page and prepare a review draft. LightOn's model card documents transcription and grounding modes, but it does not give the OCR model browser tools. A Reddit OCR challenge shows why the handoff needs a human check: even capable readers can miss strikethrough text or distort tables. LightOnOCR-3 model card · original OCR challenge · OCR evaluation discussion on Hacker News
Key takeaways
LightOnOCR-3 0.8B reads an image input and returns text. It does not plan browser steps, discover skills, or invoke Tabbit CLI.
The model card describes transcription mode for page text and grounding mode for labeled bounding boxes. Treat both as extraction, not authorization.
Use a saved page image or a locally rendered PDF page. Do not ask OCR to open a site or handle login.
A separate local agent, such as Codex with the Tabbit skill, can compare the extraction with the live page and prepare a named draft.
Human review is essential for amounts, IDs, dates, table order, strikethroughs, and low-quality scans.
Decide what each tool is allowed to do
| Component | Suitable job | Not its job |
|---|---|---|
| LightOnOCR-3 0.8B | Extract visible text or grounded page blocks from a local image. | Browse, authenticate, choose records, or click buttons. |
| Human or local agent | Select the source file, correct OCR errors, and decide what record to inspect. | Treat OCR output as verified site data. |
| Tabbit CLI in a local agent | Open the signed-in Tabbit profile, read the matching page, and prepare a draft. | Send or save changes without review. |
LightOnOCR-3 is released under Apache 2.0 according to the model card. The card describes the 0.8B version as the lightest variant and documents a Transformers image-text-to-text pipeline. The grounding prompt adds labels and bounding boxes; an empty prompt returns page text as Markdown. Use the official instructions for your installed Transformers version and verify your local hardware before processing. Model card and usage examples · LightOn's OCR repository
Read a locally saved page image
First save a page screenshot or render one PDF page to an image on your computer. Check that the file contains only data you are allowed to process. Then run the OCR model locally using the model card's Transformers example. For a page where layout matters, ask for grounding mode. Keep OCR output separate from the source file so you can compare them side by side.
Use this prompt with the OCR model:
Read only the attached local page image. Do not infer text that is not visible.
Return the visible record ID, date, status, amount, and table row labels.
Preserve the original order. Mark uncertain characters as [unclear] and missing fields as [not visible].
If the page has a strike-through, keep both the struck and replacement text and label them.
Return the approximate page region or bounding box for each extracted field when available.
Do not browse the web, use tools, log in, or suggest browser actions.Expected output is a short transcription with uncertain items labeled, not a claim that the page is accurate. A Reddit user comparing OCR tools found that a model missed strikethrough and another damaged table layout; a separate Hacker News discussion warned about hallucinated or missing text. These reports are not evaluations of LightOnOCR-3, but they support a manual comparison step. Reddit OCR challenge · Hacker News OCR discussion
Hand the result to a Tabbit-enabled local agent
Open the same record in the intended Tabbit profile and sign in yourself. If using Codex, open /skills and confirm that it has discovered the tabbit skill. The skill picker is separate from local command access. Follow the Tabbit CLI setup guide, then check that the installed launcher can reach the expected Tabbit instance. For other clients, use their documented skill syntax; Pi uses /skill:tabbit, while OpenCode documents @tabbit. Codex setup · Pi skill syntax · OpenCode skills
Give the local agent the OCR text as a clue, not as truth:
Using the Tabbit skill, open the signed-in [approved site] and inspect only record [ID from OCR].
Do not use the OCR output as verified data. Compare each field with the visible page.
Do not click edit, save, submit, approve, or send.
Return the exact ID, visible date, status, amount, and source URL. Mark differences from OCR.
Prepare a browser draft named "OCR cross-check — [date]" with the comparison and uncertain fields.
Stop before saving or sending. Show me the draft and source page for review.The expected result is a comparison table with OCR value, live page value, and a match or mismatch note. If the ID does not match, do not let the agent open a nearby record. Follow the first browser task checklist and login-session reuse guide. A browser session may expire even when the selected Tabbit profile is right. Connecting an agent to an existing browser explains the profile boundary.
Review before acting
| Field | Compare | Keep manual review when |
|---|---|---|
| Record ID | Every character against the live page URL and heading. | OCR confuses similar characters or the ID is absent. |
| Date and amount | Exact punctuation, currency, and struck-through values. | The image is low contrast or the number affects a decision. |
| Table order | Row labels and values stay on the same row. | Columns are merged, rotated, or visually separated. |
| Draft | Each proposed sentence points to the live page. | Any value comes only from OCR or model inference. |
LightOnOCR-3 0.8B is useful when a local image or scanned page needs transcription before a person or separate agent works with the live site. It is the wrong tool for browser planning or clicks. Keep the handoff explicit: local image to OCR, OCR text to a human or local agent, and only then a read-only Tabbit task with a visible review draft. For the broader architecture, read how agents control browsers, CLI versus browser MCP, and the complete CLI guide. For task separation details, see the guide to profiles, workspaces, and tab groups.
FAQ
Can LightOnOCR-3 0.8B operate a signed-in browser?
No. It reads image input and returns text; it does not browse, log in, plan browser steps, or call Tabbit CLI.
How should I use OCR output with Tabbit?
Give the extracted text to a separate local agent as a clue. Ask it to verify the matching live record in Tabbit and prepare a comparison draft.
Should OCR text be treated as verified data?
No. Check IDs, dates, amounts, table order, and uncertain characters against the original image and the live page.
Can the OCR model itself invoke the Tabbit skill?
No browser-tool integration is established for this OCR model. The user or a separate local agent must perform the handoff.
What is a safe first task?
Transcribe one local screenshot or PDF page, then have the Tabbit-enabled agent read one matching record without editing or sending anything.