Cohere's Model Vault Now Hosts Open-Weight AI Models for Businesses

Many organizations want to use AI on sensitive material, such as customer records, medical files, or internal contracts, but they can't send that material to a shared service they don't control. Cohere's latest announcement is aimed at exactly that problem.

What was announced

Model Vault is Cohere's dedicated inference platform. An inference is the moment an AI model reads your request and produces an answer, so an inference platform is the place where that work happens. Cohere describes Model Vault as fully managed, meaning the company runs the infrastructure so the customer doesn't have to build and maintain its own servers.

The new part is that Model Vault now also hosts open-weight models from other providers. An open-weight model is one whose trained settings are published, so developers can download and run it. Cohere names DeepSeek V4.1 Flash, Gemma 4, GPT-OSS, and Phi as examples, and says customers can use its own models next to these.

Two levels of protection

Cohere offers two versions of the platform:

• Model Vault gives each organization an isolated instance that serves it alone. You choose the region, and Cohere says your prompts and outputs stay there.
• Model Vault Encrypted goes further. It uses confidential computing, a technique that keeps data encrypted even while the model is processing it. Cohere says that under this setup, neither Cohere nor any cloud provider can see what is sent to a model or what comes back.

Cohere also says that zero data retention is the default in Model Vault. That means prompts and outputs are not kept after they are processed. Not every model is offered at both levels. According to the table in the announcement, Cohere's Command model is available in standard Model Vault but not in the encrypted version.

Availability is limited by region. Cohere says Model Vault is live in the United States, the European Union, and Canada. Model Vault Encrypted is live in the United States, with support for the EU and Canada described as coming soon.

What "sovereign inference" means

Cohere uses the term sovereign inference to describe control over how AI runs. The company breaks it into four parts:

1. Where your data lives (data residency, which is the physical location where information is stored).
2. Who can access it.
3. Which models you use, so you pick the best tool for each job.
4. How your AI runs, rather than having the provider decide.

The idea is that a company should not have to give up control to get access to good models.

Wrap-up

Cohere's announcement is a business update: Model Vault can now run open-weight models from several providers, with privacy controls and a choice of regions. Its claims about security and data handling are Cohere's own, so verify them with your provider. For a next step, look at the official Cohere announcement and write down the questions about data location and access your organization needs answered.