I gave OpenAI Codex and several local AI models the same messy spreadsheet job: turn three sales Excel files into an executive workbook.
Codex finished best of course. It built the workbook and flagged the key issue: no ProductID in the sales files, so exact SKU-level margin was not possible.
The local models were mixed. qwen-Coder (19GB) got stuck. qwen3.6 (23GB) got surprisingly far but needed review. It would have saved about 80% of human time. qwen3-next (50GB), though much larger, did less well than qwen3.6.
Why this test? Because some company/client data should not go to a cloud model, depending on the situation. If that is the constraint, local AI is worth testing but you have to test it against the actual job, not a generic prompt. And bigger isn’t necessarily better.
https://t.co/1sgp4ymF7D
Tested GPT-OSS-20B vs Qwen3.6 locally on a 96GB Mac Studio.
Same prompts. Same machine. Real assistant tasks.
GPT-OSS was fast.
Qwen was deeper.
That’s not a leaderboard; that’s the agent lesson:
the right system routes the right task to the right model.