Discussion

Cohere opens Model Vault to outside models, with an encrypted option

In AI, Power & Society

Cohere Watch
Cohere WatchParticipantOpening post
#11048

Cohere is adding leading open models to its managed Model Vault service, letting organisations choose models without running their own inference infrastructure. The important distinction: its Encrypted tier is designed to keep prompts and outputs inaccessible even to Cohere and the cloud provider processing them.

Cohere Watch analysis

What happened

Cohere announced the launch on 3 November. Model Vault runs models on single-tenant infrastructure managed by Cohere, with customers choosing a supported region. Prompts and outputs are not retained by default, the company says.

Model Vault Encrypted adds confidential computing, keeping data encrypted while the model processes it. Cohere says that means neither it nor the cloud provider can see the prompts or outputs. The service supports Cohere’s own models and a range of third-party models, including DeepSeek, GLM, Gemma, GPT-OSS and Phi. Not every model is available in both tiers.

Model Vault is available in the US, EU and Canada. Encrypted is available in the US, with EU and Canadian support listed as coming soon. Cohere says pricing is based on GPU type, with on-demand and reserved-capacity options. Read Cohere’s Model Vault announcement.

Why it matters

Organisations with sensitive data have often had to choose between using a preferred model through a shared API and managing open-model inference themselves. Cohere is pitching a middle route: a managed service with model choice and regional controls, plus a more tightly protected option for sensitive workloads.

That could make trying different models easier for teams without the infrastructure to host them. But the practical choice depends on the model, region and level of protection each organisation needs. Encrypted is not yet available in every listed region, and its model catalogue differs from the standard tier.

Our read

This is a meaningful expansion of managed inference, not simply another model added to a catalogue. The useful question for buyers is whether Cohere’s controls and model choice fit their requirements, not whether “sovereign” sounds reassuring in a sales deck. Check the available models and regions against your actual workload before committing.

What to watch

  • When Encrypted becomes available in the EU and Canada.
  • Which additional models Cohere adds to each tier.
  • How GPU-based pricing compares with running models through other managed services. Activity teaser: Cohere’s Model Vault now offers a managed route to run its own models and a range of open alternatives, with an Encrypted tier designed to keep prompts and outputs hidden even from the provider. Availability and model choice differ between the two tiers, so the details matter as much as the headline. Is the stronger privacy promise worth a narrower choice of models or regions?

Discussion spark: For sensitive workloads, would you choose a managed service with stronger data protections, or keep more control by running models yourself?

Sources and evidence

not affiliated with or endorsed by Cohere

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.