Views
No views yet
baidu/Unlimited-OCR,
a 3.3B-parameter vision-language model for document OCR, layout parsing, and
grounding (SAM ViT-B + CLIP ViT-L encoders → linear projector → DeepSeek-V2 MoE
LLM). Runs natively on Apple Silicon — no PyTorch or CUDA at inference time.convert.py
from the upstream bfloat16 checkpoint: MoE experts stacked to
[n_experts, out, in] for batched gather_mm, Conv2d weights transposed
OIHW → OHWI, stored in float16.1git clone https://github.com/vignesh-kumar-v/mlx-ocr-unlimited
2cd mlx-ocr-unlimited && pip install -r requirements.txt
3
4# Single image
5python ocr.py --model Vignesh-5756/Unlimited-OCR-mlx-fp16 --image doc.png --prompt "<image>Free OCR."
6
7# Multi-page PDF -> markdown
8python ocr.py --model Vignesh-5756/Unlimited-OCR-mlx-fp16 --pdf paper.pdf --output out.md--dtype fp32. Decodes at ~175 tok/s (single image) / ~120 tok/s (multi-page)
on an M5 Pro, ~6 GB resident in fp16.PreTrainedTokenizerFast (the project's loader does
this): AutoTokenizer misroutes this checkpoint to LlamaTokenizer and
corrupts the byte-level BPE.norm_topk_prob, which must default to False.