Skip to content

News

Feature

Numes adds end-to-end encrypted models to its AI inference marketplace

Numes says prompts on its new encrypted models are encrypted on the device and decrypted only inside the GPU. Independent verification is still missing.

On Nov. 13, 2026, the AI inference marketplace Numes announced on X that end-to-end encrypted models are now available. According to the company, a prompt is encrypted on the user's device and decrypted only inside the GPU that runs the model, "not even we can read it".

How it works (per the report). The approach builds on Equivariant Encryption from Nesa Research. It does not use hardware Trusted Execution Environments, which providers such as Venice or Phala rely on. Nesa's documentation claims under 9% latency overhead and over 99.99% output fidelity. Glitchwire notes that independent verification remains sparse.

What it means for users. Numes now offers two tiers: standard inference, where the provider can see prompts during processing, and encrypted inference. Numes did not say whether the encrypted models cost more or are limited in availability. Treat the privacy claim as a vendor statement until a privacy policy or an audit confirms it. Numes is not currently listed on our site, and we have not verified its payment options.

Sources