---
library_name: transformers
pipeline_tag: image-text-to-text
base_model:
- openbmb/MiniCPM-OCR-1.0
---

This is a randomly initialized tiny model based on [openbmb/MiniCPM-OCR-1.0](https://huggingface.co/openbmb/MiniCPM-OCR-1.0) for testing and debugging.

| Feature | Source | This repo |
|---|---|---|
| Checkpoint tensor elements | 1.30B BF16 | 7.58M BF16 |
| Safetensors weight files | 2.60 GB (model.safetensors) | 15.18 MB (model.safetensors) |
| Architecture / input | MiniCPM-OCR; text and document image | Same |
| Vocabulary rows / tied embeddings | 248,100 / tied | Same |
| Backbone layers / layer cadence | 24 / six 3-linear + 1-full cycles | 8 / two complete cycles |
| Hidden width / MLP width | 1,024 / 3,584 | 16 / 32 |
| Q heads / KV heads / head dimension | 8 / 2 / 256 | Same |
| Linear-attention key/value heads and dimensions | 16/16, 128/128 | Same |
| Vision layers / width / insert merger | 27 / 1,152 / after layer 6 | 4 / 128 / after layer 2 |
| Visual downsampling | 4x | Same |
| M-RoPE | sections 11/11/10; partial 0.25 | sections 11/11/10; partial 0.25 |
| MTP | disabled | Same |
| Checkpoint precision | BF16 | Same |

The tokenizer embedding and output head stay tied, so the checkpoint contains one BF16 token table. Two complete linear/linear/linear/full cycles are retained. The Qwen3.5 linear-attention path requires CUDA and Triton. The exact public snippet was verified on an NVIDIA A800-SXM4-80GB.

| File path | Size |
|------|------|
| model.safetensors | 15.2MB |

### Example usage

#### Transformers (✅ verified)

The Qwen3.5 linear-attention path requires CUDA and Triton. Install `transformers[torch]>=5.21.0` and `Pillow`.

```python
def run_transformers(model_id: str):
    import torch
    from PIL import Image, ImageDraw
    from transformers import AutoModelForImageTextToText, AutoProcessor

    model_id = str(model_id)
    processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
    model = AutoModelForImageTextToText.from_pretrained(
        model_id,
        dtype=torch.bfloat16,
        device_map="cuda",
        attn_implementation="sdpa",
    )
    image = Image.new("RGB", (448, 448), "white")
    ImageDraw.Draw(image).text((20, 20), "Hello OCR", fill="black")
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": image},
                {"type": "text", "text": "Read this document."},
            ],
        }
    ]
    inputs = processor.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
        downsample_mode="4x",
        max_slice_nums=1,
        use_image_id=False,
    ).to(model.device)
    with torch.no_grad():
        logits = model(**inputs).logits
        assert torch.isfinite(logits).all()
        output = model.generate(**inputs, downsample_mode="4x", do_sample=False, max_new_tokens=10)
    print(processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True))


if __name__ == "__main__":
    run_transformers(model_id="tiny-random/minicpm-ocr-1.0")
```

### Creating this repo

- Code: [.meta/create_model.py](.meta/create_model.py)
- Model structure: [.meta/model.log](.meta/model.log)
- Environment: [.meta/env.json](.meta/env.json)

### Change log

- 2026-11-01: Initial tiny MiniCPM-OCR 1.0 checkpoint.

*This repo is generated by DeepSeek-V4.1-Flash.*