On November 3, 2026, Cohere added third-party open models to Model Vault and Model Vault Encrypted. You can now run DeepSeek, GLM, Gemma, GPT-OSS, Phi, Qwen and others on dedicated GPUs in the US, EU or Canada, with zero data retention by default and OpenAI-compatible API calls.
Last updated: November 7, 2026
Table of contents
- What changed on November 3
- Standard vs Encrypted Vault
- Models you can deploy
- Regions and residency
- Create a vault and call open models
- GPU pricing
- Compared with shared APIs
- Power-user tips
- Frequently asked questions
- Sources
What changed on November 3
Cohere’s November 3, 2026 product post and matching changelog state that Model Vault and Model Vault Encrypted now run third-party open models beside Cohere’s own models. Regulated teams can pick an open model for a task without operating their own serving stack.
Zero data retention remains the default: prompts and responses are processed for inference and not retained afterward. Standard Vaults serve the United States, European Union and Canada. Encrypted Vaults are live in the United States, with EU and Canada listed as coming soon.
This pairs with our Cohere Rerank 5 guide. For other managed runtimes, see Fireworks Ultra-fast and Mistral Managed Agents.
Standard vs Encrypted Vault
Choose the protection level before you create the vault. The type cannot change later.
| Capability | Standard Vault | Encrypted Vault |
|---|---|---|
| Isolated tenant | Yes | Yes |
| Data in transit / at rest | Protected | Protected |
| Data in use (confidential compute) | No | Yes |
| Zero data retention default | Yes | Yes |
| Open models | DeepSeek, GLM, MiMo, Nemotron, Gemma, GPT-OSS, Phi, Qwen, Mistral families | Gemma, GPT-OSS, Phi, Qwen, Ministral, Devstral |
| Regions today | US, EU, Canada | US (EU, Canada coming soon) |
| Typical fit | Residency and isolation | Cohere must not see prompts or outputs |
Encrypted Vault docs describe hardware TEEs (for example Intel TDX or AMD SEV-SNP) plus an NVIDIA GPU in confidential-computing mode. Encrypted open models currently run on one H100 80GB confidential GPU.
Models you can deploy
Per the Standard Vault supported models page, Cohere families include Command, North (including North Mini Code), Embed, Rerank, Parse, Compass, Transcribe and customized models. Embed and Rerank are generally self-serve; generative, bundle, speech and customized models may need a waitlist.
Open-model families on Standard Vaults: DeepSeek, GLM (Z.ai), MiMo (Xiaomi), Nemotron (NVIDIA), Gemma (Google), GPT-OSS (OpenAI), Phi (Microsoft), Qwen (Alibaba) and Ministral / Devstral / Leanstral (Mistral AI). Cohere’s blog cites examples such as DeepSeek V4.1 Flash, GLM 5.3 Flash, MiMo V2.6 Pro, Gemma 4 and NVIDIA Nemotron. Exact IDs and GPUs appear in the catalogue for your region. Open models keep their developers’ licenses; review each before deploy.
Regions and residency
The regions page maps Global to the United States, EU to the European Union and CAN to Canada. Region is fixed at creation. The United States lists every supported model and GPU; the EU lists B200 or L4 models; Canada lists GB200 or L4 models (GB200 via sales). The catalogue only shows models that fit the region you picked.
Create a vault and call open models
- In the Model Vault home, click New Vault +.
- Set name, description, type (Standard or Encrypted) and region (Global, EU or CAN). Name, type and region cannot change later.
- In the catalogue, filter by provider, check context length and hourly price, pick a GPU when options exist, set replica min/max, then Deploy.
- Copy the vault endpoint URL and model name from the model card.
- For open models, point an OpenAI-compatible client at
<YOUR_VAULT_ENDPOINT_URL>/v1with your Cohere API key. Setmodelto the card name (example:deepseek-v4-1-flash). - For Cohere models on a Standard Vault, use the Cohere SDK or HTTP
/v2/chatagainst the vault endpoint, or OpenAI-compatible<YOUR_VAULT_ENDPOINT_URL>/compatibility/v1.
Example open-model call from Cohere’s docs:
from openai import OpenAI
client = OpenAI(
api_key="<COHERE_API_KEY>",
base_url="<YOUR_VAULT_ENDPOINT_URL>/v1",
)
response = client.chat.completions.create(
model="deepseek-v4-1-flash",
messages=[{"role": "user", "content": "Hello from Model Vault!"}],
)
print(response.choices[0].message.content)Organizations currently face a three-vault limit; ask Cohere for an increase if needed.
GPU pricing
Billing is per GPU per hour, times GPUs per replica, times replica count. Cohere’s pricing page (checked November 7, 2026) lists on-demand and promotional rates through March 31, 2027, including L4 24GB at $3.00 (promo $1.00), A100 40GB/80GB at $3.25/$3.50 (promo $2.25), MI300X at $6.00 (promo $4.00), H100 80GB rows at $6.25 and $9.00 (promo $4.25), H200 at $7.25 (promo $5.50), B200 rows at $9.75 and $12.00 (promo $8.00), and GB200 at $12.50 (promo $10.00). Encrypted confidential-compute GPUs use their own rates. Pricing does not guarantee availability. Idle replicas still cost GPU hours, so set minimums carefully. For pay-per-search Cohere work, see Rerank 5; for DeepSeek’s own tools, see the Search and Fetch API guide.
Compared with shared APIs
| Need | Shared public API | Model Vault |
|---|---|---|
| Any hot open model quickly | Often easiest | Depends on region and vault type |
| Single-tenant GPUs | Usually no | Yes |
| US / EU / Canada residency | Varies | Standard: all three |
| Confidential compute in use | Rare | Encrypted Vault |
| Billing | Usually tokens | GPU-hour × replicas |
| OpenAI SDK for open models | Common | Yes at vault /v1 |
Choose Vault when compliance, residency or confidential compute matter more than the lowest token price.
Power-user tips
Confirm the model string on the vault card before shipping. Open models reject Cohere /v2 and /compatibility/v1; only /v1 Chat Completions works for them. Keep Cohere models on the Cohere SDK or compatibility URL so Embed and Rerank can share the vault. Plan capacity around replica min/max. For Encrypted Vaults, budget H100 confidential pricing and the shorter open-model list. Watch the three-vault organization limit.
Related reading: Cohere Rerank 5, Fireworks Ultra-fast, Mistral Managed Agents, DeepSeek Search API and Inkling 2 Small.
Frequently asked questions
What is Cohere Model Vault?
Model Vault is Cohere’s dedicated, fully managed inference platform. Your models run on isolated infrastructure that Cohere operates. As of November 3, 2026, a Standard Vault can host Cohere models and third-party open models such as DeepSeek, GLM, MiMo, Gemma, GPT-OSS, Phi, Qwen and Mistral families.
Which open models can I run in Model Vault?
Standard Vaults support open-model families from DeepSeek, Z.ai (GLM), Xiaomi (MiMo), NVIDIA (Nemotron), Google (Gemma), OpenAI (GPT-OSS), Microsoft (Phi), Alibaba (Qwen) and Mistral AI (Ministral, Devstral, Leanstral). Encrypted Vaults support a shorter list that currently includes Gemma, GPT-OSS, Phi, Qwen, Ministral and Devstral.
What is the difference between Model Vault and Model Vault Encrypted?
A Standard Vault isolates your tenant and protects data in transit and at rest, with zero data retention by default. Model Vault Encrypted adds confidential computing so prompts and outputs stay encrypted while the model runs; Cohere states that not even Cohere or the cloud provider can see them. Encrypted Vaults currently run in the United States.
How do I call open models on Model Vault?
Open models only accept the OpenAI-compatible Chat Completions API at your vault endpoint URL plus /v1. Authenticate with your Cohere API key and set model to the name shown on your vault’s model card (for example deepseek-v4-1-flash). They do not accept Cohere /v2 API or /compatibility/v1 requests.
How much does Model Vault cost?
Pricing is per GPU per hour, multiplied by GPUs per replica and replica count. On-demand examples on Cohere’s pricing page include L4 24GB at $3.00/hour (promotional $1.00 through March 31, 2027) and H100 80GB at $6.25 or $9.00/hour depending on the row shown. Reserved capacity and commit plans are available through sales.
Which regions does Model Vault support?
Standard Vaults run in the United States (shown as Global), the European Union (EU) and Canada (CAN). Encrypted Vaults are available in the United States, with EU and Canada listed as coming soon. A vault’s region cannot be changed after creation.
Image: Cohere logo, CC0 by Ayub MK09 via Wikimedia Commons. FreeAnswers is not affiliated with Cohere.
Sources
- Cohere: Leading Open AI Models Now Available in Model Vault (Nov 3, 2026)
- Changelog: Open Models and Regional Serving in Model Vault
- Standard Vault supported models and Encrypted Vault supported models
- Calling a Standard Vault over the API and Creating a Vault
- Model Vault regions and Cohere pricing