What happened
In a November 3, 2026 post, Cohere announced third-party models on Model Vault and Model Vault Encrypted. The company describes both as fully managed inference platforms: workloads run on single-tenant infrastructure, and customers choose a supported region. Cohere says prompts and outputs are not retained by default. Its stated goal is to let organizations use different models for different tasks without operating their own inference clusters or sending sensitive material to a shared API.
The launch adds models from DeepSeek, Z.ai, Xiaomi, Google, NVIDIA, OpenAI, Microsoft and Tencent to Cohere’s own catalog. The post names DeepSeek V4.1 Flash, GLM 5.3 Flash, MiMo V2.6 Pro, Gemma 4 and NVIDIA Nemotron among the open models, and says GPT-OSS and Phi models are available too. A compatibility table shows that model availability differs by product: DeepSeek, GLM, MiMo, Nemotron and Hy3 are listed for standard Model Vault only; Gemma, GPT-OSS and Phi are listed for both Model Vault and Model Vault Encrypted. Among Cohere’s product families, Command is listed for standard Model Vault, while Embed, Rerank, Parse, Transcribe and Translate are listed for both.
The two services offer different protections. Model Vault provides an isolated instance dedicated to an organization, with customer-selected regional serving and Cohere-managed infrastructure. Model Vault Encrypted is intended for more sensitive workloads: Cohere says prompts and outputs remain encrypted while the model processes them, inside isolated environments using confidential computing. The company says neither Cohere nor the cloud provider can see the data in that mode. Those are Cohere’s product claims; the announcement does not describe independent testing or an assurance report.
Availability is not uniform. Standard Model Vault is live in the United States, European Union and Canada. Model Vault Encrypted is live in the United States; EU and Canadian regional serving is described as coming soon. Cohere says it plans to add regions, larger models in the encrypted service and more models across regions over the coming months, without giving dates. The post says pricing is based on GPU type, with on-demand hourly pricing and promotions on Cohere’s pricing page; reserved capacity is available through sales with monthly and annual commitments. It provides no price figures.
What the archive adds
Earlier Field Notes covered Cohere’s confidential-computing launch for Model Vault, including Verifiable Confidential AI Inference and Confidential Computing on Cohere Model Vault. Those notes concern the same product and establish the prior security angle. This is a new fact on that known story: the catalog now includes models from other providers, subject to model- and region-specific limits.
Trend signal
This supports agent-ready software exposes capabilities only weakly: the announcement describes a managed inference boundary with selectable models and deployment controls, but does not detail a composable agent interface or permissioned capabilities. The more defensible signal is that enterprise AI platforms are competing on control over model choice, data location and processing rather than only on a proprietary model catalog.
Why now
Cohere frames the launch as a response to organizations that want the model best suited to a task but cannot send sensitive data to shared APIs or manage open-model inference themselves. Its stated next steps—more regions, larger encrypted models and additional models—also indicate an effort to extend the service’s coverage, though the company gives no delivery dates.
Why it matters
For teams in regulated or data-sensitive settings, model selection and deployment constraints are often coupled: the model that performs best for a task may not be usable if its serving path violates residency or access requirements. Cohere’s offer could reduce the operational burden of hosting multiple open models while retaining a single managed platform and, for selected models, confidential computing.
The practical value depends on the details behind each workload. Teams need to check the compatibility table, serving region, GPU-based price and whether standard isolation is sufficient or encryption during processing is required. The announcement broadens an option; it does not establish that every model, region, workload size or compliance requirement is covered.
Limits and uncertainty
The source is Cohere’s product announcement, not an independent evaluation. It gives no latency, throughput, capacity, reliability, or comparative cost figures, and its pricing description does not include rates. It does not explain technical attestation, key management, audit evidence, or how customers can verify the confidential-computing claims. Model and region support varies, and EU and Canada availability for Model Vault Encrypted is only described as coming soon. The post also does not specify dates for the promised larger models, new regions or expanded catalog.
Reader takeaway
Model Vault’s catalog now spans Cohere and several third-party model families, but the useful unit of evaluation is not the headline list: it is the exact model, region, protection mode and GPU price your workload can actually use. Validate those against your requirements before treating the service as a substitute for a shared API or self-hosted inference.
Open question
What independent technical evidence and pricing will Cohere provide to help customers compare the encrypted service with other deployment options?