I'm considering Tabnine Pro for my team, but I'm concerned about privacy. The website says the Pro plan offers local models, but I'm not entirely sure what that means in practice.
When I use the completions, does my code get transmitted to Tabnine's servers at all? Even if it's just for processing? I handle client data sometimes, so I need to be clear on where the code goes.
"Local models" usually means the model weights are on your machine. The inference runs locally.
But check your network traffic. A lot of these tools phone home for "metrics" or "model updates" constantly. The completion might be local, but your IDE plugin could be sending snippets to verify context or gather usage data.
If you handle client data, assume it's leaving your network until you prove otherwise with a packet sniffer. Their privacy policy will have more weasel words than a politician's speech.
If it ain't broke, don't 'upgrade' it.
Good point on the network traffic. The "model updates" loop is a known data exfiltration trick.
Run `lsof -i` or check your egress logs. You'll see the connections.
- Initial download of the model weights (obvious, one-time).
- Constant, smaller HTTPS calls to their API every 15 minutes.
Those small calls contain telemetry. Might be "anonymized," but it's still your file paths and surrounding code context leaving your machine. Their privacy policy calls it "product improvement data."
You didn't really get local inference, you got a data collection client.
show the math