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, AutoModelForCausalLM, set_seed

    folder = Path(save_folder)
    folder.mkdir(parents=True, exist_ok=True)
    # LFM3 is not registered in the installed Transformers release; ship its pinned
    # remote code so the checkpoint loads without a newer package.
    for name in (
        "configuration_lfm3.py",
        "modeling_lfm3.py",
        "lfm3_conv.py",
        "chat_template.jinja",
        "generation_config.json",
        "special_tokens_map.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())
    # Keep the two-conv prefix and two complete full/conv/sliding/conv cycles.
    # Match tiny-random/lfm2's residual/MLP widths; retain LFM3's 4:1 Q/KV ratio.
    config.update(
        hidden_size=64,
        intermediate_size=128,
        num_hidden_layers=10,
        num_attention_heads=4,
        num_key_value_heads=1,
        head_dim=32,
        max_position_embeddings=128000,
    )
    config["layer_types"] = config["layer_types"][:10]
    (folder / "config.json").write_text(json.dumps(config, indent=2) + "\n")

    set_seed(42)
    model = AutoModelForCausalLM.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/LFM3-1.7B",
    source_revision="ac2586a34f1a1c03329dbdcdefabd8b6478ddaa7",
    save_folder="/tmp/lfm3",
)