Cohere Model Vault adds DeepSeek, GLM, Gemma and other open models
Single-tenant, zero-retention hosting for open models, billed by the GPU hour.
Cohere added third-party open models to Model Vault, its single-tenant managed inference platform, on November 3, 2026. Enterprises can now run DeepSeek, GLM, MiMo, Gemma, Nemotron, GPT-OSS, Phi and Hy3 models next to Cohere's own models, with prompts and outputs not retained by default.
What changed
- Model Vault now serves open model families from DeepSeek, Z.ai, Xiaomi, Google, NVIDIA, OpenAI, Microsoft and Tencent. Cohere names DeepSeek V4.1 Flash, GLM 5.3 Flash, MiMo V2.6 Pro, Gemma 4 and NVIDIA Nemotron.
- Each customer runs on single-tenant infrastructure in a region it chooses, managed by Cohere. Zero data retention of prompts and outputs is the default.
- Model Vault Encrypted keeps data encrypted while the model processes it, using confidential computing. Of the open models, Gemma, GPT-OSS and Phi run there today.
- Model Vault is live in the United States, the European Union and Canada. Model Vault Encrypted is live in the US, with the EU and Canada coming soon.
- Pricing is per GPU hour, by GPU type and number of replicas, not per token. Cohere lists on-demand and promotional hourly rates, the promotions running through March 31, 2027, and sells reserved capacity on monthly and annual commits.
Open models by tier
From Cohere's announcement:
| Model family | Developer | Model Vault | Model Vault Encrypted |
|---|---|---|---|
| DeepSeek | DeepSeek | Yes | No |
| GLM | Z.ai | Yes | No |
| MiMo | Xiaomi | Yes | No |
| Gemma | Yes | Yes | |
| Nemotron | NVIDIA | Yes | No |
| GPT-OSS | OpenAI | Yes | Yes |
| Phi | Microsoft | Yes | Yes |
| Hy3 | Tencent | Yes | No |
What it means
Banks, hospitals and public-sector teams that cannot send data to a shared API get a managed, single-tenant way to run widely used open models, without building their own cluster. Because Model Vault bills by the GPU hour, the cost per token depends on how much traffic you push through each replica. Per-token APIs remain the simpler fit for low or uneven volume.
The same models on RouterPlus
Cohere is not a provider in the RouterPlus catalog. Lowest input and lowest output prices for the same models on shared per-token endpoints in the catalog, November 3, 2026, per million tokens:
| Model | Lowest input | Lowest output |
|---|---|---|
| DeepSeek V4.1 Flash | $0.12 | $0.42 |
| DeepSeek V4.1 Flash 1017 | $0.09 | $0.30 |
| GLM 5.3 Flash | $0.05 | $0.20 |
| MiMo V2.6 Pro | $0.387 | $0.783 |
| Gemma 4 31B | $0.12 | $0.35 |
| gpt-oss-120b | $0.037 | $0.17 |
| Nemotron 3.5 Super VL | $0.15 | $0.50 |
| Hy3 | $0.14 | $0.58 |
Compare every host on the RouterPlus models page. If you need dedicated capacity instead of shared endpoints, tell us what you need.