Repository navigation
v0.16.0: native Unlimited-OCR
Native Swift + MLX runtime for Unlimited-OCR (DeepSeek-OCR). krill run unlimited-ocr --image <page> "document parsing." parses documents and images to grounded text on Apple Silicon — no Python, no trust_remote_code. A DeepSeek-MoE language backbone + the native DeepEncoder vision tower (SAM-ViT-B + CLIP-L + projector), with base-view vision features spliced at the <image> block before the LM. Reads real multi-line invoices (titles, line-item tables, totals) correctly.
Ships as a 2.2 GB mixed-precision Krill blob (srv-sngh/Unlimited-OCR-mixed-nvfp4, from 6.67 GB bf16): MoE experts at nvfp4, attention / FFN / embed / lm_head / vision at 8-bit affine. nvfp4 expert support is an additive mode on the shared switched-expert runtime, so every other MoE family is unchanged.
krill pull unlimited-ocr
krill run unlimited-ocr --image page.png "document parsing."
Serves the parity-validated base view (full pages incl. wide layouts); gundam tiling for very dense scans is a tracked follow-up. Built on baidu/Unlimited-OCR (MIT) — credit to the original authors.