Model Vault is Cohere’s managed inference service for running its models on dedicated serving infrastructure. Standard Vault provides logical isolation and encryption in transit and at rest; Encrypted Vault adds hardware-backed protection while data is processed and client-verifiable attestation. The encrypted offering is a design-partner beta with its own model list and access conditions. Applications can call vault endpoints directly, or run North in their own network while Cohere operates the inference layer.
Managed model inference and confidential computing
What it enables
Core features
Configurable performance tiers and 1–25 replicas per model, with request, latency, token and GPU monitoring
Standard Vault isolates serving components but shares authentication and underlying Kubernetes/compute resources; it does not provide encryption in use
Encrypted Vault combines confidential CPU virtual machines and attested NVIDIA GPUs, with OHTTP encryption into the protected environment
Local OHTTP proxy or conseel Python transport verifies attestation before sending requests; the hosted demonstration proxy introduces additional trust in Cohere
Zero Data Retention applies when enabled; OHTTP keys remain inside the attested environment
Standard Fixed plans reserve instances; Flex adds autoscaling and instance-hour overages against agreed latency objectives
Encrypted beta lists Embed, Rerank, Parse, Compass, North Mini Code, Transcribe and customized models; confirm the exact version, GPU and region with Cohere
Defence and security uses
Managed inference for enterprise search, document processing and supervised agent workflows
Confidential-computing evaluation for sensitive inference workloads, subject to workload-specific assurance
Evidence of maturity
Standard Vault is commercially offered; Encrypted Vault is documented as a design-partner beta. Published architecture describes vendor controls rather than independent workload acceptance. No attained TRL is stated.
Commercial availability: Standard Embed/Rerank are self-serve; other models may require a waitlist. Encrypted beta is free for design partners through sales enrollment; general-availability pricing and Canadian region availability require confirmation.
OHTTP private keys are generated inside the TEE; attestation-gated peer sharing allows replicas to decrypt requests. Cohere states prompts/responses are not retained when ZDR is enabled; this is a configuration-dependent claim.
Supports: Capability · technical tags
OHTTP private keys are generated inside the TEE; attestation-gated peer sharing allows replicas to decrypt requests. Cohere states prompts/responses are not retained when ZDR is enabled; this is a configuration-dependent claim.
Cohere states Encrypted Vault processes inference inside confidential CPU VMs and attested NVIDIA GPUs, with client OHTTP encryption terminating inside the TEE. Hardware/firmware, client and attestation remain trusted; side channels, traffic metadata and operator-controlled availability remain residual risks.
Supports: Capability · summary
Cohere states Encrypted Vault processes inference inside confidential CPU VMs and attested NVIDIA GPUs, with client OHTTP encryption terminating inside the TEE. Hardware/firmware, client and attestation remain trusted; side channels, traffic metadata and operator-controlled availability remain residual risks.
Supports: Capability · technicaldomainslugs
Cohere states Encrypted Vault processes inference inside confidential CPU VMs and attested NVIDIA GPUs, with client OHTTP encryption terminating inside the TEE. Hardware/firmware, client and attestation remain trusted; side channels, traffic metadata and operator-controlled availability remain residual risks.
Supports: Capability · defence applications
Cohere states Encrypted Vault processes inference inside confidential CPU VMs and attested NVIDIA GPUs, with client OHTTP encryption terminating inside the TEE. Hardware/firmware, client and attestation remain trusted; side channels, traffic metadata and operator-controlled availability remain residual risks.
03
Model Vault Overview
Cohere
Source details
Source type: official company product
Supports: Capability · technical tags
Cohere describes Model Vault as a managed single-tenant inference service with dedicated model-serving components, configurable model/performance tiers, 1–25 replicas per model and monitoring; it can serve applications directly or North.
Supports: Capability · capability type
Cohere describes Model Vault as a managed single-tenant inference service with dedicated model-serving components, configurable model/performance tiers, 1–25 replicas per model and monitoring; it can serve applications directly or North.
Supports: Capability · summary
Cohere describes Model Vault as a managed single-tenant inference service with dedicated model-serving components, configurable model/performance tiers, 1–25 replicas per model and monitoring; it can serve applications directly or North.
Supports: Capability · core features
Cohere describes Model Vault as a managed single-tenant inference service with dedicated model-serving components, configurable model/performance tiers, 1–25 replicas per model and monitoring; it can serve applications directly or North.
Supports: Capability · name
Cohere describes Model Vault as a managed single-tenant inference service with dedicated model-serving components, configurable model/performance tiers, 1–25 replicas per model and monitoring; it can serve applications directly or North.
04
Encrypted Vault Supported Models
Cohere
Source details
Source type: official company product
Supports: Capability · core features
Encrypted Vault beta lists Embed, Rerank, Parse, Compass, North Mini Code, Transcribe and customized models. Model, GPU, capacity and region availability require dashboard or sales confirmation; the list does not establish Command family parity with Standard Vault.
Supports: Capability · commercial availability
Encrypted Vault beta lists Embed, Rerank, Parse, Compass, North Mini Code, Transcribe and customized models. Model, GPU, capacity and region availability require dashboard or sales confirmation; the list does not establish Command family parity with Standard Vault.
Cohere documents Encrypted Vault as beta, free for design partners, with sales-led enrollment and full pricing to be shared as it matures. This does not establish generally available commercial encrypted service.
Supports: Capability · commercial availability
Cohere documents Encrypted Vault as beta, free for design partners, with sales-led enrollment and full pricing to be shared as it matures. This does not establish generally available commercial encrypted service.
Supports: Capability · technology readiness level
Cohere documents Encrypted Vault as beta, free for design partners, with sales-led enrollment and full pricing to be shared as it matures. This does not establish generally available commercial encrypted service.
Supports: Capability · maturity
Cohere documents Encrypted Vault as beta, free for design partners, with sales-led enrollment and full pricing to be shared as it matures. This does not establish generally available commercial encrypted service.
Standard Vault offers Fixed reserved instances or Flex baseline plus autoscaling with additional instance-hour billing, monthly/annual commitments and pause/resume. Tier selection follows latency/throughput SLOs; model/tier sizing matters more than a general no-rate-limit promise.
Supports: Capability · maturity
Standard Vault offers Fixed reserved instances or Flex baseline plus autoscaling with additional instance-hour billing, monthly/annual commitments and pause/resume. Tier selection follows latency/throughput SLOs; model/tier sizing matters more than a general no-rate-limit promise.
Standard Vault protects data in transit and at rest without hardware protection of data in use or remote attestation. Its architecture shares authentication and underlying Kubernetes/compute resources; Zero Data Retention applies when enabled.
Supports: Capability · core features
Standard Vault protects data in transit and at rest without hardware protection of data in use or remote attestation. Its architecture shares authentication and underlying Kubernetes/compute resources; Zero Data Retention applies when enabled.
A local OHTTP proxy container or conseel Python transport checks attestation before encrypting requests. The optional hosted demo proxy changes the trust model by requiring trust in Cohere at the proxy layer and is disabled by default.
Standard Vault lists Command, North, Embed, Rerank, Parse, Compass, Transcribe and customized models. Embed/Rerank are self-serve; other models or bundles may need a waitlist. GPU and region availability are confirmed with Cohere.
Supports: Capability · defence applications
Standard Vault lists Command, North, Embed, Rerank, Parse, Compass, Transcribe and customized models. Embed/Rerank are self-serve; other models or bundles may need a waitlist. GPU and region availability are confirmed with Cohere.
Intel Trust Authority checks composite CPU/GPU evidence and registered policy; the client verifies its signed token and bound OHTTP key before sending data. Attestation identifies code, not whether that code is free of vulnerabilities. Cohere documents reproducible builds and remaining hardware, firmware and client trust.
Customers run North application/control-plane components in their VPC/network and route inference to Cohere-managed vault endpoints. Standard and Encrypted types work with North, with attestation verification before encrypted inference.
Supports: Capability · summary
Customers run North application/control-plane components in their VPC/network and route inference to Cohere-managed vault endpoints. Standard and Encrypted types work with North, with attestation verification before encrypted inference.
Supports: Capability · technicaldomainslugs
Customers run North application/control-plane components in their VPC/network and route inference to Cohere-managed vault endpoints. Standard and Encrypted types work with North, with attestation verification before encrypted inference.
Cohere states Encrypted Vault processes inference inside confidential CPU VMs and attested NVIDIA GPUs, with client OHTTP encryption terminating inside the TEE. Hardware/firmware, client and attestation remain trusted; side channels, traffic metadata and operator-controlled availability remain residual risks.
Cohere describes Model Vault as a managed single-tenant inference service with dedicated model-serving components, configurable model/performance tiers, 1–25 replicas per model and monitoring; it can serve applications directly or North.
