def create_model(source_model_id: str, source_revision: str, save_folder: str) -> str:
    import json
    from pathlib import Path

    import torch
    from huggingface_hub import hf_hub_download
    from transformers import AutoConfig, AutoModelForImageTextToText, set_seed

    folder = Path(save_folder)
    folder.mkdir(parents=True, exist_ok=True)
    # LFM3-VL is not registered in the installed Transformers release; ship its pinned remote code (text backbone,
    # vision tower and the image/video processors) so the checkpoint loads without a newer package.
    for name in (
        "configuration_lfm3.py",
        "configuration_lfm3_vl.py",
        "modeling_lfm3.py",
        "modeling_lfm3_vl.py",
        "lfm3_conv.py",
        "processing_lfm3_vl.py",
        "image_processing_lfm3_vl.py",
        "video_processing_lfm3_vl.py",
        "chat_template.jinja",
        "generation_config.json",
        "processor_config.json",
        "tokenizer.json",
        "tokenizer_config.json",
        "LICENSE",
    ):
        (folder / name).write_bytes(Path(hf_hub_download(source_model_id, name, revision=source_revision)).read_bytes())

    config = json.loads(Path(hf_hub_download(source_model_id, "config.json", revision=source_revision)).read_text())
    # Text backbone: residual/MLP widths 64/128 like tiny-random/lfm3, with at least four KV heads for TP=4 and
    # the source 4:1 Q/KV ratio (16 Q heads, 4 KV heads, rectangular projections at head dim 32), the two-conv
    # prefix plus two complete full/conv/sliding/conv cycles.
    text_config = config["text_config"]
    text_config.update(
        hidden_size=64,
        intermediate_size=128,
        num_hidden_layers=10,
        num_attention_heads=16,
        num_key_value_heads=4,
        head_dim=32,
        max_position_embeddings=128000,
    )
    text_config["layer_types"] = text_config["layer_types"][:10]
    # Vision tower: at least four KV heads with the source 2:1 Q/KV ratio (8 Q heads, 4 KV heads). Its head dim is
    # derived as hidden // num_heads, so width 256 keeps head dim 32 and axial RoPE; patch 16 / temporal patch 2 /
    # spatial merge 2. Six layers retain the conv/conv prefix and two full/conv cycles.
    vision_config = config["vision_config"]
    vision_config.update(
        hidden_size=256,
        intermediate_size=640,
        num_hidden_layers=6,
        num_attention_heads=8,
        num_key_value_heads=4,
    )
    vision_config["layer_types"] = ["conv", "conv", "full_attention", "conv", "full_attention", "conv"]
    # Projector: LayerNorm + GELU MLP merging 2x2 vision blocks; keep bias and the downsample contract.
    config["projector_hidden_size"] = 128
    (folder / "config.json").write_text(json.dumps(config, indent=2) + "\n")

    set_seed(42)
    model = AutoModelForImageTextToText.from_config(
        AutoConfig.from_pretrained(folder, trust_remote_code=True), trust_remote_code=True, dtype=torch.bfloat16
    )
    model.save_pretrained(folder)
    # `save_pretrained()` rewrites these two files with this environment's versions; restore the pinned source bytes
    # plus the deliberate config edits so only the intended fields differ from the source.
    (folder / "config.json").write_text(json.dumps(config, indent=2) + "\n")
    (folder / "generation_config.json").write_bytes(
        Path(hf_hub_download(source_model_id, "generation_config.json", revision=source_revision)).read_bytes()
    )
    return str(model)


create_model(
    source_model_id="LiquidAI/LFM3-VL-2B",
    source_revision="ba44bd51db20a2753f28c273a1571ccaa210d954",
    save_folder="/tmp/lfm3-vl",
)