I announced open source OCR with VLMs last week that costs $0.14 per 1k pages. We tripled…

I announced open source OCR with VLMs last week that costs $0.14 per 1k pages. We tripled throughput and cut cost by 3x since then. Today, we’re releasing all the details and the code. We processed… | Mathis Lucka

I announced open source OCR with VLMs last week that costs $0.14 per 1k pages. We tripled throughput and cut cost by 3x since then. Today, we’re releasing all the details and the code. We processed 116,000 PDF pages on a single L40S GPU using Modal as our compute provider. The full production run took 89 minutes and cost $4.64. That’s $0.04 per 1k pages at a speed of 20 pages per second. It’s 100x cheaper than Mistral OCR 4 and 250x cheaper than Azure’s document intelligence. We use GLM-OCR as the VLM in combination with PP-DocLayoutV3 for layout detection. This setup has SOTA scores on OmniDocBench and excels at complex documents and tables. Click through to our blog (comments) to learn everything about the system, how we optimized it, and to access the full code base.

lnkd.in