Cohere opens Model Vault to rival models on dedicated infrastructure

Cohere expanded Model Vault on November 3 to run third-party open models alongside its own offerings. The release notice describes fully managed, dedicated infrastructure with zero data retention for prompts and responses by default.
The standard service now supports models from labs including DeepSeek, Alibaba, Google, OpenAI, Microsoft, NVIDIA and Mistral. Customers call outside models through an OpenAI-compatible Chat Completions interface at their vault endpoint.
For teams weighing hosted APIs against self-managed inference, the offering provides another division of responsibility: the customer chooses the model and capacity, while Cohere operates the serving infrastructure.
Model choice depends on the protection tier
The encrypted tier has a narrower catalog. Cohere’s announcement lists Gemma, GPT-OSS, Phi, Qwen, Devstral and Ministral for confidential-computing deployments. Standard Vault runs in the United States, European Union and Canada; Encrypted Vault is available in the United States, with the other two regions described as forthcoming.
That makes model, region and protection level a combined purchasing decision. A model supported somewhere in the service is not necessarily supported in the location and tier a customer requires.
Cohere’s service overview distinguishes dedicated infrastructure and encryption in transit and at rest from protection during computation. The latter, plus verifiable remote attestation, belongs to Encrypted Vault.
The encrypted-service documentation says requests pass through an OHTTP proxy that checks the deployment’s attestation and encrypts requests before they leave the client machine. Cohere says the design prevents it and the cloud provider from viewing prompts and responses.
Those are specific service claims, not a blanket guarantee that an application using the service meets every privacy or compliance obligation. An application may still write sensitive material to its own logs, monitoring system or downstream tools.
Capacity economics replace a simple token comparison
Cohere prices both tiers per GPU per hour. Buyers therefore need to consider utilization and required capacity, rather than compare the offering only against a provider’s per-token rate.
A steady workload may use dedicated capacity differently from an application that sits idle for long periods and then experiences short bursts. Evaluations should include throughput, queueing and the cost of keeping enough capacity available.
Model Vault’s expansion gives open-model users a managed deployment option with a wider catalog. The practical shortlist still begins with three checks: the exact model, the supported region and the level of protection the workload needs.
Sources and citations
Revision history
Published .