Jina AI built its newest document parser to run on hardware most teams already own.
The target is the NVIDIA L4, a modest inference card. jina-ocr-v1, released by Jina AI, now part of Elastic, reads PDFs, scans, charts, tables and invoices and returns Markdown in one pass.
Its size explains the pitch. The checkpoint totals 3.4B parameters, but only about 570M decoder parameters are active per token. A speculative decoding head ships inside the same file, so no separate draft model has to be served.
Accuracy numbers come from two public suites: 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench.
The model post-trains DeepSeek-OCR. Encoder duty falls to a DeepEncoder of roughly 380M parameters. It chains three known pieces together: SAM, a 16x convolutional compressor, and CLIP-L.
Compression is aggressive. One 1024×1024 page view arrives as 4,096 patches and leaves as 256 visual tokens. A dynamic mode adds up to nine local tiles, holding a page near 1,156 tokens. Decoding runs through a 12-layer mixture-of-experts block with top-6 routing.
Speed comes from a FastMTP head, which drafts three tokens recursively. The decoder verifies them greedily and commits the longest matching prefix. Output therefore equals plain greedy decoding, so the acceleration costs no quality. About 2.73 tokens land per step.
The checkpoint lands near 6.8 GB in BF16. Transformers and vLLM both load it. Licensing is CC BY-NC 4.0, which covers research and non-commercial work, while anyone shipping it commercially has to get a license from Jina AI.
