Cohere Vault Adds Open Models and Rerank 5 for Enterprise Retrieval

Cohere updated its documentation on 3 November 2026 to confirm that Model Vault and Model Vault Encrypted now run leading third-party open models alongside its proprietary systems.reported This expansion allows organisations to select specific models for distinct tasks while maintaining consistent security protections across their infrastructure.This update lets companies run popular open-source AI models on secure, dedicated servers managed by Cohere.The changelog confirms multi-vendor open model support in Model Vault alongside the release of Rerank 5 with context-aware retrieval capabilities.

The evidence

The changelog states that Standard Vault supports models from DeepSeek, Z.ai, Xiaomi, NVIDIA, Google, OpenAI, Microsoft, Alibaba and Mistral AI. Encrypted Vault currently supports Gemma, GPT-OSS, Phi, Qwen, Devstral and Ministral. These encrypted deployments use confidential computing to ensure that neither Cohere nor the cloud provider can view prompts or responses.

The decision

Regional serving options differ between the two vault types. Model Vault operates in the United States, the European Union and Canada. Model Vault Encrypted is currently available only in the United States, with European Union and Canadian regions described as coming soon. Pricing is structured per GPU per hour for both standard and encrypted tiers.

What the evidence means

Developers interact with these open models via an OpenAI-compatible Chat Completions API at their vault endpoint. The system enforces zero data retention by default, meaning prompts and responses are not stored. Users seeking reserved capacity or private deployments in their own VPC or on-premises are directed to contact the sales team.

What builders can do with it

Builders can integrate Rerank 5, released on 21 October 2026, into their retrieval pipelines via the Rerank API, Microsoft Foundry or Amazon SageMaker.confirmed The pro variant targets high-quality ranking for complex reasoning tasks, while the fast variant prioritises low latency and high throughput. Both variants support a 32k token context window and over 100 languages.

Builders can utilise the Context Focus feature to receive character spans of the most relevant passages within documents. This allows applications to pass only those snippets to large language models, thereby reducing token costs or highlighting matches for users. Calibrated relevance scores enable consistent thresholds across different queries and domains.