Skip to navigation
Model VaultStandard Vault

Supported Models

A Standard Vault runs both Cohere’s models and leading third-party open models. For the deployment flow, see Creating a Vault.

Cohere models

Model familyType
CommandGenerative, including reasoning, multimodal, and translation variants
NorthGenerative models, including North Mini Code
EmbedEmbeddings
RerankReranker
ParseVision parser
CompassEmbed + Rerank + Parse bundle
Cohere TranscribeSpeech recognition (ASR), including Arabic
Customized Cohere modelsVaries

Embed and Rerank models are generally available self-serve. Generative models, bundles, speech, and customized models may require a waitlist. For the specific versions you can deploy, check the model catalogue when creating a vault, see the models overview, or contact Cohere.

Open models

A Standard Vault runs third-party open models from the following families. When you create a vault, the model catalogue shows the specific models and GPU options available in the region you selected. Use the model name shown on your vault’s details page in API requests.

Model familyDeveloper
DeepSeekDeepSeek
GLMZ.ai
MiMoXiaomi
NemotronNVIDIA
GemmaGoogle
GPT-OSSOpenAI
PhiMicrosoft
QwenAlibaba
Ministral, Devstral, LeanstralMistral AI

Open models are called through the OpenAI-compatible Chat Completions API, not the Cohere SDK. See Calling open models.

Open models are provided under their developers’ licenses. Review the license for each model on its developer’s model page before you deploy it. Specific model availability may change over time.

Supported GPUs

A Standard Vault runs on NVIDIA L4, A100, H100, B200, and GB200 GPUs and AMD MI300X GPUs. When a model runs on more than one GPU, you choose the GPU when you add it to a vault, and the GPU determines the regions the model is available in. NVIDIA GB200 GPUs are available through Cohere sales. To confirm the options available for your model, capacity, and region, contact Cohere.

Differences for Encrypted Vaults

Model and GPU availability differs for Encrypted Vaults, since confidential computing depends on specific GPU hardware.

Model Vault is priced per GPU, per hour: you pay the hourly rate for each GPU your model runs on, multiplied by the number of GPUs per replica and the number of replicas. See the pricing page.