model : add LightOnOCR-1B model

[ngxson]

Seems like the "OCR model race" has started. This seems to be one of the few "low hanging fruits" that we can easily support in llama.cpp

The model features:

• Qwen3 as language model
• Mistral3 as vision encoder (the difference is that LightOnOCR does not use [IMG_BREAK] token)

Original model: https://huggingface.co/lightonai/LightOnOCR-1B-1025

GGUF model: https://huggingface.co/ggml-org/LightOnOCR-1B-1025-GGUF

To try it:

llama-cli -hf ggml-org/LightOnOCR-1B-1025-GGUF -c 8192

# open https://localhost:8080 and try uploading an image

---

Important note: this model requires specific input structure, see the chat template

The structure seems to be:

• Starts with an empty system message
• Then, an user message. All images must be contained in this message; No instructions are needed

Example:

{
  "messages": [{
    "role": "system",
    "content": ""
  }, {
    "role": "user",
    "content": [{
      "type": "image_url",
      "image_url": {"url": "data:image/png;base64,......"}
    }]
  }],
}

[ggerganov]

Very cool!

The command in OP should be llama-server instead of llama-cli.