How Do I Use Private (Confidential) AI Models?

Bankr can run certain AI models inside a Trusted Execution Environment (TEE) — hardware-secured enclaves where your prompts and the model's responses are processed so that even the infrastructure provider can't see the data. This is called private (confidential) inference.

How to Use It

Append :private to any supported model name — for example bankr/glm-4:private instead of bankr/glm-4. That's the whole switch, and it works anywhere you pick a model (web terminal, CLI, API).

Bankr uses a "fail-closed" policy: if you request a :private model and a secure enclave isn't available, the request fails with an error rather than silently falling back to a non-private connection. Your data is never accidentally exposed.

Which Models Support It

Private inference is available for open-weight models — families like DeepSeek, GLM (Z.ai), Kimi (Moonshot), MiniMax, and Gemma. Proprietary models (Claude, GPT, Gemini) don't support confidential serving through this method.

For the current list, run bankr llm models --private in the CLI, or check the model lineup at bankr.bot/llm.

Good to Know

• Responses are verifiable. Private responses include attestation headers proving the computation happened inside a verified enclave.

• Error 422 (confidential_unavailable) — you added :private to a model that doesn't support TEE serving. Pick an open-weight model instead.

• Error 503 (attestation_unverified) — the gateway couldn't verify the hardware's security certificate, so it blocked the request to keep your data safe. Try again.

For API examples, Claude Code configuration, and the attestation verification flow, see Private Inference on docs.bankr.bot.