solus_v1_unlimited-ocr-q5

Unlimited OCR 3B (Q5_K_M) — Solus v1

baidu's Unlimited-OCR converted to GGUF: a SAM + CLIP "DeepEncoder" vision tower feeding a DeepSeek MoE text decoder, for multilingual OCR and document parsing. It reads a page image and returns markdown, with grounding markers that Solus strips before showing the text. Built on the DeepSeek-OCR stack, but its own model: a ring sliding-window KV cache keeps the full image prefill resident while rotating only recent decode tokens, which is what lets it run long documents without unbounded KV growth.

This is the Q5_K_M build, the more accurate of the two readers in the Solus catalogue. It ships with a multimodal projector (mmproj) that must be loaded alongside the weights; without it the model loads but sees nothing. Needs a llama.cpp with DeepSeek-OCR support (merged upstream in PR #17400); the projector declares clip.projector_type = deepseekocr and the weights general.architecture = deepseek2-ocr.

Specifications

Single file: Unlimited-OCR-Q5_K_M.gguf

Quantization

Quantization performed at the Faculty of Engineering, McMaster University.

The conversion this build is derived from was produced by sahilchachra, and the weights here are a byte-for-byte copy of that file — the SHA-256 above matches the upstream artifact.

Provenance

• Original model: baidu/Unlimited-OCR
• Upstream GGUF: sahilchachra/Unlimited-OCR-GGUF
• Mirrored for Solus, a desktop app for running language models entirely on your own machine.

Usage

llama-cli -m Unlimited-OCR-Q5_K_M.gguf -cnv

License

Licensed MIT.