Model Vault is Cohere’s dedicated model-inference platform; the company describes the beta in an October 22, 2026 technical post.
According to Cohere, the serving stack runs in a confidential virtual machine and NVIDIA Confidential Computing extends protection to GPU memory while encrypting CPU-to-GPU traffic. Request contents are additionally protected with Oblivious HTTP (OHTTP): the payload is encrypted to a public key associated with the attested environment, and Cohere says only an OHTTP gateway inside that environment can decrypt it.
The client evaluates signed remote-attestation evidence against an approved policy before releasing a protected request. Cohere says a failed check, such as changed software measurements or a GPU not in the required confidential mode, stops the client from sending the payload. The post does not provide pricing, availability regions, performance figures, supported models, or an independent security audit, and it does not claim to remove all trusted components from the system.