groxaxo/TeleOCR-oQ8-MLX

TeleOCR oMLX oQ8

Text-and-vision OCR checkpoint quantized on this Mac from the official TeleOCR BF16 weights. The base model is XingChen-AGI/TeleOCR revision e92585356c0d0b7b7a65938f3da035c6593cc9a6. Upstream license is Apache-2.0. Project page: caipeng328/TeleOCR.

What this file is

model.safetensors is 2,125,686,530 bytes. config.json records affine quantization, 8 bits, group size 64. SHA256SUMS is the hash of the uploaded files.

The quantizer was oMLX quantize_oq_streaming from checkout jundot/omlx commit 70f44cfd652d2013dc0f9776dca7518e32101b00, with oq_level=8, group_size=64, BF16 residuals, text_only=False, enhanced=False, and trust_remote_code=True. Calibration used the built-in code_multilingual corpus, 128 samples of 256 tokens. This level finished in 9.683 seconds on 7 October 2026. The quantizer logged that the mlx-vlm sanitize chain preserves the vision weights. It recorded layer-sensitivity scores of 0 and did not print a bits-per-weight plan line.

MLX reported a peak of 3,546,131,464 bytes during the three-level run that produced oQ8, then oQ6, then oQ4. The counter was not reset between levels.

teleocr_mlx.py is the local adapter. TeleOCR uses 128-wide heads, bias-free projections, and per-head query/key RMSNorm. model_file in config.json points at that adapter. This folder was not run through the 2 October OCR comparison. The measured character-error figures for that comparison belong to groxaxo/TeleOCR-oQ5-MLX and the BF16 source, not to this 8-bit file.

The pinned BF16 source SHA-256 recorded on 2 October 2026 is 9817b18041bd96403f75e38a33b28ed0cd5fb2641f67eead673afabd3c408109.

Use

Serve the parent directory with an oMLX build that can load the bundled teleocr_mlx.py adapter. The directory name is the model id.

omlx serve --model-dir /path/to/parent --host 127.0.0.1 --port 8000

Request model TeleOCR-oQ8.