Today, I'm not testing chat—I'm letting my local model get to work. After running Ling-3.0 Tiny locally, I’ve always felt that chatting alone isn’t very meaningful. What’s truly worth testing is whether it can be integrated into an actual workflow. So this time, instead of redeploying the model, I directly used llama.cpp’s local API to build the simplest possible “AI Text Organizer.” My setup: Windows + llama.cpp + Ling-3.0 Tiny, with the model serving an API at 127.0.0.1:8080. First, confirm the model is running: http://127.0.0.1:8080/v1/models If you see model information, the API is working. Then create a new file called ai_test.py and paste in the following code: import requests API_URL = "http://127.0.0.1:8080/v1/chat/completions" text = """ Today's meeting mainly discussed the launch of the new product. The UI is nearly complete, but there are two bugs in the payment module. The product manager wants testing completed by Friday. The backend team needs to recheck the order status API. Also, the promotional copy hasn’t been finalized yet—the designer needs to update the homepage based on the final version. """ prompt = f""" You are a work information organizer. Please organize the following messy work notes into: 1. One-sentence summary 2. Main issues 3. To-do items 4. Deadlines 5. Tasks requiring others' cooperation Requirements: - Do not invent information not present in the original text - If information is uncertain, write “unspecified” - Use Chinese - Output clearly, but avoid overcomplicating Original content: {text} """ data = { "model": "local-model", "messages": [ { "role": "user", "content": prompt } ], "temperature": 0.2, "max_tokens": 1000 } response = requests.post(API_URL, json=data) if response.status_code == 200: result = response.json() print(result["choices"][0]["message"]["content"]) else: print("Request failed:", response.status_code) print(response.text) Then install Python’s requests: pip install requests Finally, run: python ai_test.py Now the model isn’t just “chatting with me”—it’s actively processing real work notes. I then replaced the test content with something even more disorganized: This project has been messy lately—the homepage is almost done, but login occasionally throws errors; Xiao Wang said he fixed it yesterday, payment hasn’t been fully tested yet, ideally we should have results by Thursday, the promotional graphics are still being revised, and the boss says it must go live by month-end, but the exact date hasn’t been set yet. This kind of test is far more meaningful than asking the model “Who are you?” Because it deliberately includes: completed tasks, incomplete tasks, vague timelines, different people’s responsibilities, and undefined deadlines. What matters most is whether the model can distinguish facts from assumptions. If it correctly identifies “must go live by month-end” as a deadline, while not turning “ideally have results by Thursday” into a formal deadline, this small test has succeeded. The next step doesn’t even require changing the model—just replace the text with: meeting transcripts, customer service chats, group messages, project requirements, or personal notes. Even a tiny 100MB local model can now enter real workflows. This is what I find most interesting about local AI. It’s not about making it a “local ChatGPT”—it’s about turning it into a small AI tool inside your computer that never uploads raw data and can be directly called by programs. And this system can be extended further: read TXT files, batch-process dozens of documents, integrate with web interfaces, or even build a simple desktop UI. This is much closer to what I understand as “local AI truly being put into practice” than simply testing whether the model can chat.