unlimited-ocr-max

baidu/Unlimited-OCR served through MAX as an OpenAI-compatible endpoint on Metal, CUDA, ROCm or CPU. Weights: kthierbach/unlimited-ocr-max.

• bf16 — baidu's model.safetensors, unchanged.
• int8 — model-int8.safetensors, this port's weight-only quantisation of the routed experts (symmetric, per-group G=128).
• config.json is baidu's minus auto_map and model_type, so MAX loads it without remote code.

Prerequisites

Python 3.12 or 3.13, plus one of:

Apple silicon — M1–M5, macOS 15+, 24 GB recommended

# requires full Xcode (App Store); the Command Line Tools alone are not enough
sudo xcode-select -s /Applications/Xcode.app/Contents/Developer
# MAX compiles Metal kernels with `xcrun metallib`, shipped in the Metal Toolchain
xcodebuild -downloadComponent MetalToolchain
xcrun -f metallib          # must print a path

NVIDIA — Linux (glibc 2.34+, e.g. Ubuntu 22.04+), Ampere or newer, driver 580+

Follow NVIDIA's package-manager installation; for example, on Ubuntu 24.04 x86_64:

# MAX loads these at runtime but does not ship them; without them the
# first request fails with: symbol not found: cublasCreate_v2
# 1. add NVIDIA's CUDA repository (ubuntu2404/x86_64 shown; use your distro/arch)
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
# 2. libcublas + libcublasLt (one package), CUDA 13
sudo apt-get install -y libcublas-13-0
# 3. check: MAX loads these exact files
ls -l /usr/local/cuda-13.0/lib64/libcublas.so.13 /usr/local/cuda-13.0/lib64/libcublasLt.so.13

Use these system packages, not the nvidia-*-cu13 pip wheels — the loader does not find those. NVIDIA needs v0.3.1 or later: v0.3.0 aborts on the first request with CUDA_ERROR_INVALID_VALUE.

AMD — Linux, driver 6.3.3+ (MI355X: ROCm 7+)

MAX also loads ROCm's rocBLAS, hipBLASLt and MIOpen from /opt/rocm/lib; untested beyond compilation.

CPU

Nothing beyond Python. Supported, slow.

Serve

uv tool install unlimited-ocr-max        # or: pip install unlimited-ocr-max
unlimited-ocr-max serve --devices gpu

This installs max[all]==26.6.0, numpy and pillow from PyPI (no extra index). The first run downloads the weights (6.2 GiB bf16, 4.0 GiB int8) and compiles the kernels. Endpoint: http://127.0.0.1:8010/v1/chat/completions, model unlimited-ocr-max.

Measure it on your machine

unlimited-ocr-max profile --devices gpu            # add --weights int8, or --devices cpu

Starts its own server, sends the 12 bundled pages, prints one row of the table below and writes profile.json. On a GPU it refuses to run unless no other process is using the GPU. --out DIR sets the output directory (default ./unlimited-ocr-max-profile-<UTC time>).

One page to Markdown:

curl -s http://127.0.0.1:8010/v1/chat/completions -H 'Content-Type: application/json' -d @- <<EOF
{"model": "unlimited-ocr-max", "temperature": 0, "max_tokens": 1766,
 "messages": [{"role": "user", "content": [
   {"type": "text", "text": "<|grounding|>Convert the document to markdown."},
   {"type": "image_url", "image_url": {"url": "data:image/png;base64,$(base64 < page.png | tr -d '\n')"}}]}]}
EOF

Offline:

uvx --from huggingface_hub hf download kthierbach/unlimited-ocr-max --revision v0.3.1 --local-dir ocr-model
unlimited-ocr-max serve --devices gpu --model ocr-model

Where it has run

temperature 0, default guard. Text is compared against the fp32 PyTorch reference (transformers 4.46.3, CPU); CER = edited characters / reference characters over all pages.

The three Apple M4 rows are one draw each with unlimited-ocr-max profile on v0.3.1 (macOS 26.5.2, max 26.6.0); the other rows predate profile and were not re-run. Peak memory is the first request's compile; steady is the server idle after the last page. ¹ Upstream: MAX's ldmatrix PTX needs sm_80, and Turing has no bf16 tensor cores (modular/modular#6653, #6659). ² System swap grew during these runs, so their timings are indicative; the text results are exact.

Not supported

gundam (tiled) mode; batch size > 1; multi-GPU.

References

Unlimited OCR Works (Yin et al., 2026), arXiv:2606.23050. This port changes how the model is served, not the model (int8 excepted, as quantified above); model-level behaviour, limitations and biases are those of baidu/Unlimited-OCR.

Authorship

This port, its serve configuration and this documentation were written with substantial AI assistance (Claude, via Claude Code); Konstantin Thierbach reviewed them and is accountable for their contents. Assisted-by: AI.

License

MIT — see LICENSE: this port's MIT notice and baidu's verbatim. The bf16 weights, tokenizer files and config.json are baidu's, from baidu/Unlimited-OCR (MIT, Copyright (c) 2026 Baidu), redistributed under that license; the bf16 weights are unchanged, config.json has two keys removed, and model-int8.safetensors is derived from those weights by this port's quantiser. The profile corpus bundles renders of four pages of that paper (arXiv:2606.23050); see unlimited_ocr_max/profile_data/NOTICE.md.