tiny-random/minicpm-robot-manip

This is a randomly initialized tiny model based on openbmb/MiniCPM-RobotManip for testing and debugging.

Feature	Source	This repo
Checkpoint tensor elements	1.30B BF16, 235.81M F32	8.29M BF16, 459.10K F32
Safetensors weight files	3.54 GB (model.safetensors)	18.45 MB (model.safetensors)
VLM text layers / width / head dimension	24 / 1024 / 256	8 / 32 / 32
VLM attention cadence	3 linear + 1 full, repeated	Two 3-linear + 1-full cycles
Vision layers / width	27 / 1152	2 / 32
Action diffusion blocks / width	16 / 768	2 / 32
Action horizon / action-state dimensions	30 / 80-80	Same
Embodiment categories	32	Same
Text vocabulary rows	248094	Same
Stored VLM / action-head precision	BF16 / FP32	Same
Input / output	Image, instruction, state → action chunk	Same

File path	Size
model.safetensors	18.4MB

Example usage

Transformers (✅ verified)

def run_transformers(model_id: str):
    import torch
    from PIL import Image
    from transformers import AutoModel, AutoProcessor

    model = AutoModel.from_pretrained(model_id, trust_remote_code=True, dtype=torch.bfloat16).eval().cuda()
    processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
    image = Image.new("RGB", (112, 112), "blue")
    inputs = processor.apply_chat_template(
        [
            {
                "role": "user",
                "content": [{"type": "image", "image": image}, {"type": "text", "text": "Pick up the blue object."}],
            }
        ],
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
        processor_kwargs={"padding": False},
    ).to(model.device)
    action = model.predict_action(
        state=torch.zeros(1, 1, 80, device=model.device),
        embodiment_id=torch.zeros(1, dtype=torch.long, device=model.device),
        **inputs,
    )
    assert action.shape == (1, 30, 80)
    assert torch.isfinite(action).all()
    print(action.shape)


if __name__ == "__main__":
    run_transformers(model_id="tiny-random/minicpm-robot-manip")

Creating this repo

• Code: .meta/create_model.py
• Model structure: .meta/model.log
• Environment: .meta/env.json

Change log

• 2026-10-27: Created a tiny MiniCPM RobotManip vision-language-action policy.

This repo is generated by GPT-6.1-Sol.