The four deployment modes, and what each actually means
This is the reason to consider Tabnine, so it is worth being precise.
SaaS — Tabnine’s infrastructure, like any hosted competitor. No advantage over Copilot here.
Single-tenant cloud — a dedicated instance in a cloud environment, isolated from other customers. Addresses multi-tenancy concerns without requiring you to run hardware.
On-premises — inside your own network, on your servers. Code never crosses your boundary.
Air-gapped — no external connection at all, ever. This is the one almost nothing else in this category offers, and for defence, classified work and some healthcare and financial environments it is not a preference but a precondition.
Each step down that list increases what you must run and maintain. Air-gapped means GPUs you own, updates you apply manually, and staff who understand the deployment — a genuine infrastructure project rather than an extension install.
Training data provenance, and what indemnification covers
The second pillar is what the models learned from. Tabnine offers models trained on permissively licensed code, and provides indemnification on enterprise plans.
The claim being made is that generated code will not carry a licence obligation you did not know about — that a completion will not turn out to be copyleft-derived and quietly infect a proprietary codebase.
Read what the indemnity actually covers before relying on it. Indemnities differ in scope, cap, and what conduct voids them, and the useful question for your legal team is not “is there an indemnity” but “what does it pay, under what conditions, and who decides”. A directory entry cannot answer that; the contract can.
What the indemnity does do, reliably, is convert an open legal question into a contractual one. For many procurement processes that is the entire point — the risk stops being unbounded and starts being someone’s obligation.
The capability trade, stated plainly
Restricting training data costs capability. Running on your own hardware costs capability. Both are true and both are the price of what you are buying.
Tabnine’s completions are generally not as strong as the frontier-model tools, and its agentic features are behind the tools built around agents from the start. Anyone evaluating it on raw quality against Cursor or Copilot will conclude it loses, and they will be right and also measuring the wrong thing.
The honest framing: if your organisation has no compliance driver, you are paying for guarantees you do not need and accepting weaker output to get them. If it does, nothing else in this category is available to you at all.
Codebase personalisation
Tabnine can learn from your own repositories so suggestions follow your existing patterns, naming and internal APIs rather than generic ones.
This matters more in a large organisation than the raw model quality does. A completion that uses your internal HTTP client correctly is more useful than a better-written one that invents a different approach, and in a codebase with strong conventions this closes much of the gap to the frontier tools.
It also requires the deployment to see your code, which is exactly what the on-premises and air-gapped modes make acceptable.
What each deployment mode demands
- A supported editor — VS Code, JetBrains IDEs, Visual Studio, Eclipse, Neovim and others. Coverage of older enterprise IDEs is unusually good and is sometimes the deciding factor.
- For self-hosted or air-gapped deployment: real infrastructure. Servers, GPUs for the model, and someone who owns it. This is an operations project.
- An enterprise agreement for the governance features — self-hosting and indemnification are not on individual plans.
- For SaaS use, an ordinary account and an internet connection.
Who actually buys this
Regulated industries and security-conscious enterprises: banks, insurers, healthcare providers, defence contractors, public sector — anywhere code cannot leave the network as a matter of policy or law rather than preference.
Organisations whose legal function has asked a hard question about training data and will not accept “the vendor filters similar code” as an answer.
Teams on older enterprise IDEs that the newer tools do not support at all.
For an individual developer or a startup without those constraints, it is a hard sell — the governance is the product and you would be buying it unused.
What the governance actually buys
- Air-gapped operation is genuinely possible — the strongest reason to choose it, and rare.
- Licence provenance and indemnification, converting a legal objection into a contractual one.
- Broad editor support, including IDEs everyone else ignores.
- Codebase personalisation, which closes much of the quality gap in a conventional codebase.
- Administrative control over policy, access and audit.
- A procurement story that survives a security review, which is itself worth money.
What you pay for it, in capability
- Model capability trails the frontier — the deliberate trade, and a real one.
- Self-hosting is expensive in money and staff time: GPUs, maintenance, upgrades.
- Weaker agentic features than tools designed around agents.
- Enterprise sales cycle — the features that justify it are not self-serve.
- Overkill without a compliance driver, where cheaper tools do more.
- Personalisation needs your code, so the benefit and the deployment decision are coupled.
Exit
Uninstall the extensions. Your code was always yours and, in the self-hosted modes, never went anywhere.
What is worth planning for is the infrastructure: an air-gapped deployment is hardware and configuration you own, and decommissioning it is a project rather than a cancellation. Any personalisation model trained on your code is not portable to another vendor.
What else to consider
- GitHub Copilot — stronger models and enterprise controls, but no air-gapped option at all.
- Sourcegraph Cody — the other enterprise-oriented option, strong on large-codebase context, also self-hostable.
- Cline plus Ollama — a fully local assembly if you have the skills, accepting a smaller model.
- Codeium — self-hosting options with a more current feature set.
Compiled from Tabnine’s documentation and public sources. Indemnification and deployment terms are contractual — verify against your own agreement rather than any summary. We have not hands-on tested this tool. Last reviewed 16 August 2026.
