---

license: apache-2.0

language:

• pl
• en
• ru

tags:

• ocr
• vision
• image-text-to-text
• bielik
• mistral3
• gguf
• polish

pipeline_tag: image-text-to-text

library_name: gguf

---

<!-- meta-description: Bielik Custom OCR Vision — a Polish text-recognition model merged from a Mistral-3 vision encoder and Bielik-7B. Reads printed/scanned text (strong Polish diacritics), translates on request, runs fully local on 24GB Apple Silicon via AnythingLLM / OpenWebUI. Q4_K_M GGUF. -->

Bielik Custom OCR Vision — local Polish text recognition

A custom merged model: the vision layers were borrowed from Mistral-3 and joined with

Bielik-7B (SpeakLeash). The result is a compact multimodal model that reads text from images

— printed pages, scans and documents — with excellent handling of Polish and its diacritics

(ą, ć, ę, ł, ń, ó, ś, ź, ż). On request it will also **translate the recognized text into foreign

languages**, always answering in flawless language.

The model is honest about what it is. It is built for text recognition (OCR), and it is

very good at it. It is not a general image-understanding model: it was **not fine-tuned for

scene/object recognition**, because that requires a real training lab, not a hobby homelab — we

don't have the setup to train models at that scale, so we didn't pretend to. What it does, it does

well; what it wasn't trained for, it doesn't claim.

What it's for

• OCR / text transcription from images — faithful, character-for-character, layout preserved

(line breaks, paragraphs, indentation; tables as plain aligned text, no markdown).

• Strong Polish — correct diacritics, punctuation, casing, numbers, dates and amounts.
• On-request translation of the recognized text into foreign languages, in flawless target

language.

• Fully local & private — documents never leave your machine.

Measured behavior

• Faithful transcription by default — transcribes what is on the image, preserves the layout,

marks unreadable fragments as [nieczytelne], and returns an empty answer when there is no text

— no hallucinated content. (For Polish text see the note below.)

• Two clean modes — plain OCR by default; correct-Polish answers and translation only when the

user explicitly asks. The faithful transcription is never altered by the language polish.

• Excellent on real text — clean recognition of Polish documents in day-to-day use through

AnythingLLM and OpenWebUI.

Built-in Polish proofreading (a feature, not a bug)

The language half of this model is Bielik-7B — a model trained natively on Polish. Because of

that, it doesn't just copy Polish text, it understands it: on Polish input it will

automatically correct spelling and restore missing diacritics — e.g. it turns Rafal into

Rafał, zl into zł, and quietly fixes typos it recognises. This is simply what a native Polish

model does; the Polish language knowledge is baked into the weights.

Treat it as a bonus: the OCR output comes out as clean, correct Polish rather than a raw,

diacritic-stripped scan. In testing, English and Russian scans were transcribed exactly, 1:1

(no such "correction" there), while Polish scans came back polished — proper ogonki and

orthography. If you ever need a strictly literal transcription that preserves original Polish

misspellings, this is the one thing to keep in mind — but for reading and reusing documents, having

the text auto-normalised to correct Polish is a genuine plus.

Not designed for

• General image recognition (objects, scenes, charts, diagrams) — out of scope. The vision

path was merged in, not fine-tuned on broad visual data, so treat this strictly as a text

reader. Proper visual grounding would need a full training lab; this is a focused OCR tool, by

design.

Context

• 8K tokens configured (num_ctx 8192) — comfortable for single pages and typical documents.
• Underlying architecture natively supports a much larger window; for very dense multi-page scans,

raise num_ctx if your hardware has the headroom.

Test hardware

The model was built and tested on:

• Mac Mini M4, 24 GB RAM, macOS — Ollama, GPU (Metal) inference, loads 100% on GPU.
• Orchestration: AnythingLLM and OpenWebUI — both drove the model for real OCR workloads.
• Quantization: Q4_K_M GGUF (~6 GB).

How it was made

The two models were merged using Claude Code together with Fable 5, and the model was **tuned

with Claude Opus 4.8** — the best models in the world. The merge, the system prompt, the parameter

choices and the two-mode (OCR / translation) design draw directly on that work: top-tier models

preparing a small, private text reader that runs right on your desk.

Available files

| File | Quant | Size | Notes |

|---|---|---|---|

| bielik-custom-ocr-vision-Q4_K_M.gguf | Q4_K_M | ~6 GB | OCR-tuned build; runs fully on GPU on 24 GB-class Apple Silicon. |

Ships with the OCR system prompt and parameters (see Modelfile): faithful transcription by

default, correct Polish and on-request translation, temperature 0.

License

Apache 2.0 (inherited from the Mistral-3 and Bielik-7B bases).