Table of Contents
Cohere has added third-party open models to Model Vault, expanding its managed inference platform beyond its own AI models. Enterprises can now run models including DeepSeek V4.1 Flash, GLM 5.3 Flash, MiMo V2.6 Pro, and Gemma 4 on dedicated infrastructure, with zero retention of prompts and outputs by default.
The expanded catalog also includes NVIDIA Nemotron, OpenAI's GPT-OSS, Microsoft's Phi, and Tencent's Hy3. A subset is available through Model Vault Encrypted, which uses confidential computing to protect customer data during processing as well as in storage and transit.
Organizations handling sensitive information get a choice of model providers along with controls over where inference runs and who can access the data. Cohere manages the serving infrastructure, removing the need for customers to build and operate their own GPU clusters.
Dedicated Infrastructure and Confidential Computing
Inference is the process of running a trained AI model to generate a response or perform a task. Hosting it involves more than obtaining model weights. Teams also need computing capacity, serving software, and the operational resources to keep workloads running.
Model Vault supplies that stack through an isolated instance dedicated to one organization. Customers select a supported region, and their prompts and outputs remain there. Its single-tenant design separates an organization's workloads from those of other customers, while the default retention policy means prompts and responses are not kept after processing.
Model Vault Encrypted adds protection for data in use. Confidential computing processes workloads inside isolated, protected environments, addressing the stage when a model must work with customer information rather than merely store or transmit it. Cohere says neither it nor the underlying cloud provider can see prompts or outputs in the encrypted tier.
The two offerings therefore address related but different requirements. Dedicated infrastructure and regional processing provide workload isolation and data residency controls. The encrypted tier also restricts provider visibility into the information being processed. A customer choosing between them needs to match those protections to both its workload and the models supported in each tier.
Which Models Run in Each Tier
Gemma, GPT-OSS, and Phi are available in both Model Vault and Model Vault Encrypted. DeepSeek, GLM, MiMo, Nemotron, and Hy3 currently run in the standard tier only.
Cohere's own catalog has the same split. Command, its family for generative and agentic tasks, is available in Model Vault. Its embedding, reranking, document-processing, speech, and translation models are also available in the encrypted tier.
| Model Family | Developer | Model Vault | Model Vault Encrypted |
|---|---|---|---|
| Command | Cohere | Available | Not available |
| Embed, Rerank, Parse, Transcribe, and Translate | Cohere | Available | Available |
| DeepSeek | DeepSeek | Available | Not available |
| GLM | Z.ai | Available | Not available |
| MiMo | Xiaomi | Available | Not available |
| Gemma | Available | Available | |
| Nemotron | NVIDIA | Available | Not available |
| GPT-OSS | OpenAI | Available | Available |
| Phi | Microsoft | Available | Available |
| Hy3 | Tencent | Available | Not available |
The breadth of the catalog lets teams select different models for different parts of an application without separately operating each model's infrastructure. For example, a document workflow can use Parse to process documents, Embed and Rerank to support retrieval, and a generative model to produce responses. Model choice still determines whether the entire workflow can run within the encrypted tier.
Language Weaver, the translation business within RWS, is already using Model Vault to run models, including its LW Pro model. RWS research and product development executive Dragos Munteanu said the service provides data residency and zero-retention controls while Cohere handles hosting and scaling, allowing his team to focus on its translation products.
Regional Availability, Pricing, and Expansion
Model Vault is available in the United States, the European Union, and Canada. Model Vault Encrypted is currently available in the United States, with EU and Canadian serving planned.
The expansion builds on Model Vault's January launch, when the platform offered a way to run Cohere models with dedicated infrastructure and data controls. Third-party support broadens that choice while keeping inference within the same managed service.
Cohere prices Model Vault by GPU type, with on-demand hourly options listed through its pricing page. Reserved capacity is available through sales with monthly and annual commitments. Customers can access the platform through the Model Vault portal.
Over the coming months, Cohere plans to add regions, expand the model selection within each region, and bring larger models to Model Vault Encrypted. The planned EU and Canadian encrypted deployments would extend the confidential-computing option to customers requiring processing in those jurisdictions.
