How LocoOperator-4B Surpassed Local AI Models?

python -m venv .venv
source .venv/bin/activate

![Screenshot from How LocoOperator-4B Surpassed Local AI Models? at 155s](/ai/locooperator/locooperator-155.webp)

pip install --upgrade pip
pip install torch transformers accelerate sentencepiece safetensors
pip install jupyter

jupyter notebook

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "locooperator-4b"  # replace with the exact repo id
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)

allocated_gb = torch.cuda.memory_allocated() / 1e9
print(f"VRAM allocated: {allocated_gb:.2f} GB")

nvidia-smi

import json
from transformers import TextStreamer

system = (
    "You are LocoOperator-4B. Your only job is to propose a single tool call as JSON. "
    "Respond with one JSON object with keys: name, arguments. No prose."
)
user = (
    "Explore the repo at /home/user/my_project and tell me what the main entry point is. "
    "Prefer Python files that start the app."
)

template = f"<|system|>\n{system}\n<|user|>\n{user}\n<|assistant|>\n"
inputs = tokenizer(template, return_tensors="pt").to(model.device)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

out = model.generate(
    **inputs,
    max_new_tokens=400,
    do_sample=False,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.eos_token_id,
)
raw = tokenizer.decode(out[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True).strip()
print(raw)

{
  "name": "search_repo",
  "arguments": {
    "path": "/home/user/my_project",
    "query": "main entry point",
    "globs": ["**/*.py"],
    "max_results": 100
  }
}

import os
import json
import glob
from pathlib import Path
import requests
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "locooperator-4b"  # replace with the exact repo id
PROJECT_PATH = "/home/user/my_project"
OPENROUTER_API_KEY = os.environ.get("OPENROUTER_API_KEY")

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)

def ask_for_tool_call(path: str, query: str) -> dict:
    system = (
        "You are LocoOperator-4B. Output exactly one JSON object with keys: name, arguments. "
        "No extra text."
    )
    user = f"Explore the repo at {path} and find the main entry point. Query: {query}"
    prompt = f"<|system|>\n{system}\n<|user|>\n{user}\n<|assistant|>\n"

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    out = model.generate(
        **inputs,
        max_new_tokens=400,
        do_sample=False,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
    )
    raw = tokenizer.decode(out[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True).strip()
    return json.loads(raw)

def execute_tool_call(tool_call: dict) -> dict:
    name = tool_call.get("name", "")
    args = tool_call.get("arguments", {})
    path = args.get("path", PROJECT_PATH)
    query = args.get("query", "main")
    globs = args.get("globs", ["**/*.py"])

    files = []
    for pattern in globs:
        files.extend(glob.glob(str(Path(path) / pattern), recursive=True))

    matches = []
    for f in files:
        try:
            text = Path(f).read_text(encoding="utf-8", errors="ignore")
        except Exception:
            continue
        if "if __name__ == '__main__'" in text or "uvicorn.run" in text or query in text:
            matches.append({"file": f, "snippet": text[:2000]})
    return {"query": query, "matches": matches[:10]}

def send_to_claude(context: dict) -> str:
    if not OPENROUTER_API_KEY:
        return "OpenRouter key not set. Skipping."
    headers = {
        "Authorization": f"Bearer {OPENROUTER_API_KEY}",
        "Content-Type": "application/json",
    }
    payload = {
        "model": "anthropic/claude-3.5-sonnet",
        "messages": [
            {"role": "system", "content": "You are a senior code reviewer."},
            {"role": "user", "content": f"Given these file matches, identify the main entry point:\n{json.dumps(context) }"}
        ],
    }
    r = requests.post("https://openrouter.ai/api/v1/chat/completions", headers=headers, json=payload, timeout=120)
    r.raise_for_status()
    data = r.json()
    return data["choices"][0]["message"]["content"]

tool_call = ask_for_tool_call(PROJECT_PATH, "main entry point")
fs_result = execute_tool_call(tool_call)
analysis = send_to_claude(fs_result)

print("Tool call:", json.dumps(tool_call, indent=2))
print("Local FS result keys:", list(fs_result.keys()))
print("Claude analysis:", analysis[:800])