Runtime and measured quantization ladder for tencent/HY-Embodied-0.5 - bf16 reference and 4/5/6/8-bit MLX weights, measured pointing accuracy.
Vimal Nakrani PRO
vimalnakrani
AI & ML interests
None yet
Recent Activity
updated a collection 1 day ago
HY-Embodied-0.5 MLX updated a collection 1 day ago
HY-Embodied-0.5 MLX updated a collection 1 day ago
HY-Embodied-0.5 MLXOrganizations
None yet
Measured CER/WER ladders vs BF16 for baidu/Unlimited-OCR: MLX and GGUF quants, incl. the BF16 references the ladders are measured against.
-
vimalnakrani/unlimited-ocr-bf16-mlx
Image-Text-to-Text • 3B • Updated • 103 • 1 -
vimalnakrani/unlimited-ocr-6bit-mlx
Image-Text-to-Text • 1B • Updated • 111 -
vimalnakrani/unlimited-ocr-8bit-mlx
Image-Text-to-Text • 1B • Updated • 83 -
vimalnakrani/unlimited-ocr-4bit-mixed-mlx
Image-Text-to-Text • 0.9B • Updated • 141
Runtime and measured quantization ladder for tencent/HY-Embodied-0.5 - bf16 reference and 4/5/6/8-bit MLX weights, measured pointing accuracy.
Measured CER/WER ladders vs BF16 for baidu/Unlimited-OCR: MLX and GGUF quants, incl. the BF16 references the ladders are measured against.
-
vimalnakrani/unlimited-ocr-bf16-mlx
Image-Text-to-Text • 3B • Updated • 103 • 1 -
vimalnakrani/unlimited-ocr-6bit-mlx
Image-Text-to-Text • 1B • Updated • 111 -
vimalnakrani/unlimited-ocr-8bit-mlx
Image-Text-to-Text • 1B • Updated • 83 -
vimalnakrani/unlimited-ocr-4bit-mixed-mlx
Image-Text-to-Text • 0.9B • Updated • 141