Document OCR

Pass Holo a document page as an image and get clean Markdown back: headings, lists, tables, and equations, in reading order. There is no dedicated OCR endpoint; it is the same OpenAI-compatible chat/completions call with an image plus a transcription prompt. Send it as a single request with temperature=0.0 and enable_thinking=False so the model transcribes in one shot instead of reasoning first.

Holo OCR is strongest on English, digitally generated documents (exported PDFs, slides, web pages, reports). Scanned pages and photos are best-effort, and handwriting is not a good fit. For high-stakes handwritten or non-Latin content, use a dedicated OCR system.

Set up the OpenAI client first by following the Quickstart.

Transcribe a page

Send one page image and read the Markdown from message.content.

IMAGE_URL = "https://your-host/page.png"  # or "data:image/png;base64,..."

OCR_PROMPT = (
    "Transcribe this document page to Markdown, preserving the reading order, "
    "headings, lists, and tables. Render tables as Markdown tables and equations "
    "as LaTeX. Return only the transcription, with no commentary and no surrounding "
    "code fence. If the page has no readable text, return an empty string."
)

response = client.chat.completions.create(
    model="holo3-1-35b-a3b",
    messages=[{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": IMAGE_URL}},
            {"type": "text", "text": OCR_PROMPT},
        ],
    }],
    temperature=0.0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)

print(response.choices[0].message.content)

const IMAGE_URL = "https://your-host/page.png"; // or "data:image/png;base64,..."

const OCR_PROMPT =
  "Transcribe this document page to Markdown, preserving the reading order, " +
  "headings, lists, and tables. Render tables as Markdown tables and equations " +
  "as LaTeX. Return only the transcription, with no commentary and no surrounding " +
  "code fence. If the page has no readable text, return an empty string.";

const response = await client.chat.completions.create({
  model: "holo3-1-35b-a3b",
  messages: [
    {
      role: "user",
      content: [
        { type: "image_url", image_url: { url: IMAGE_URL } },
        { type: "text", text: OCR_PROMPT },
      ],
    },
  ],
  temperature: 0.0,
  // chat_template_kwargs is H-specific, passed through in the request body
  ...({ chat_template_kwargs: { enable_thinking: false } } as any),
});

console.log(response.choices[0].message.content);

Multi-page PDFs

Holo reads images, not PDFs, so rasterize each page to an image and transcribe them one per request, then stitch the results. One page per request keeps each image at full resolution and is the most reliable pattern.

import base64
import pymupdf  # pip install pymupdf

def ocr_page(png_bytes: bytes) -> str:
    data_uri = "data:image/png;base64," + base64.b64encode(png_bytes).decode()
    response = client.chat.completions.create(
        model="holo3-1-35b-a3b",
        messages=[{
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": data_uri}},
                {"type": "text", "text": OCR_PROMPT},
            ],
        }],
        temperature=0.0,
        extra_body={"chat_template_kwargs": {"enable_thinking": False}},
    )
    return response.choices[0].message.content or ""

with pymupdf.open("document.pdf") as doc:
    pages = [ocr_page(page.get_pixmap(dpi=200).tobytes("png")) for page in doc]

markdown = "\n\n".join(pages)
print(markdown)

import { pdf } from "pdf-to-img"; // npm install pdf-to-img

async function ocrPage(png: Buffer): Promise<string> {
  const dataUri = "data:image/png;base64," + png.toString("base64");
  const response = await client.chat.completions.create({
    model: "holo3-1-35b-a3b",
    messages: [
      {
        role: "user",
        content: [
          { type: "image_url", image_url: { url: dataUri } },
          { type: "text", text: OCR_PROMPT },
        ],
      },
    ],
    temperature: 0.0,
    ...({ chat_template_kwargs: { enable_thinking: false } } as any),
  });
  return response.choices[0].message.content ?? "";
}

const pages: string[] = [];
for await (const page of await pdf("document.pdf", { scale: 2 })) {
  pages.push(await ocrPage(page));
}

const markdown = pages.join("\n\n");
console.log(markdown);

Rasterize at roughly 150 to 200 DPI (or scale: 2). Lower resolution loses small text; much higher wastes tokens without improving accuracy. Run pages concurrently to speed up long documents, within your rate limit.

holo3-1-35b-a3b caps output at 4,096 tokens per request, and a dense page (large tables, small print) can exceed that: check finish_reason and treat length as a truncated transcription. Split the page image, or switch to holo3-122b-a10b (32,768-token output cap) for dense documents. See Models.

Next steps

Element localization

Get click coordinates from a screenshot.

Agent loop

How to use Holo in your computer-use harness.

API reference

Endpoint, models, parameters, and limits.

Get started

Guides

Transcribe a page

Multi-page PDFs

Next steps

Element localization

Agent loop

API reference

​Transcribe a page

​Multi-page PDFs

​Next steps

Element localization

Agent loop

API reference

Transcribe a page

Multi-page PDFs

Next steps