Deference Incognito

Deference Incognito is a route on the Deference API that runs supported models on Phala's confidential-computing hardware. Users add :incognito to a model id and keep the same API key, SDK and credit. Calls use zero data retention routing and do not fall back to another provider.

What It Is

Incognito is a privacy-oriented inference route for developers and coding agents. A request such as qwen/qwen3.8-27b:incognito is carried by Deference and its routing provider to Phala, where Intel TDX and NVIDIA confidential GPUs seal model memory from the host and operator. The listed models come from Qwen, DeepSeek, GLM, Kimi, gpt-oss and Tencent Hy.

How a call is handled

• Only the model id changes; the rest of the code stays the same.
• If Phala cannot serve the request, the call fails instead of being rerouted.
• Deference states that it does not store prompt or response text. It keeps model, tokens and cost for billing.
• Deference and its routing provider do process the text in transit, and each hop runs over TLS.

Checking what served a call

The Activity view names the provider that served each Incognito request, with a status of provider reported. Phala publishes a hardware attestation for each model. It covers the enclave hardware, not an individual call.

Agent use

Coding agents can run Incognito models through Deference MCP, using the deference_run_model tool with an :incognito model id.

Current verification status

The product page and Incognito model-pricing page returned origin connection timeouts during the November 1, 2026 review. The product and security descriptions above are retained vendor claims from earlier extraction; current availability and model-specific rates could not be independently rechecked.