We've been using GitHub Copilot for about a year. Management pushed for a centralized, on-prem-friendly AI tool, so we switched to Tabnine Enterprise last week. 10 backend/devops engineers.
Initial setup was straightforward. The admin dashboard is clean. The big sell was the private code model trained on our repos.
First impressions after a few days:
* **Completions feel more conservative** than Copilot. Fewer wild suggestions, which is probably good for our codebase.
* **Kubernetes YAML and Terraform HCL support is solid.** It nails repetitive resource definitions.
* **The "per-team" model training is a mixed bag.** For our standard tooling, it's great. For obscure internal scripts, it still struggles.
* **Speed is noticeable.** Runs on our infra, so no latency issues.
Biggest gripe so far: the IDE plugins feel less polished. VS Code integration sometimes needs a restart to pick up model updates.
Cost-wise, it's comparable. We're giving it a 90-day trial. If the model gets smarter with our code, it might stick.