What happened
On 3 November 2026 Cohere published “The world’s leading open models, fully under your control”. The post says: “Today we’re launching third party models inside Model Vault and Model Vault Encrypted.” Model Vault is Cohere’s managed hosting service for running models in a dedicated environment; Model Vault Encrypted is the variant with additional encryption. Cohere says both are available today.
Key details
- Models named in the post: DeepSeek V4.1 Flash, Gemma 4, GPT-OSS and Phi, plus GLM 5.3 Flash, MiMo V2.6 Pro and NVIDIA Nemotron in the body text. Cohere’s support table also lists Hy3 from Tencent.
- What runs where: all eight models are marked as supported in Model Vault. In Model Vault Encrypted, the table marks Gemma, GPT-OSS and Phi as supported and DeepSeek, GLM, MiMo, Nemotron and Hy3 as not supported.
- Data handling: “Zero data retention of prompts and outputs is implemented by default in Model Vault,” according to Cohere.
- Regions: Model Vault has regional serving in the United States, the European Union and Canada. Model Vault Encrypted is available in the US, with the EU and Canada listed as coming soon.
- Pricing: Cohere says pricing is by GPU type, with on-demand hourly rates on its pricing page and reserved capacity through sales with monthly and annual commitments. The post gives no per-token or dollar figures.
- Not stated in the post: the license terms of the third-party models, the GPU prices, performance numbers or which model versions are served.
Why it matters
Several of the listed models come from Chinese labs, and the offer is aimed at organisations that want to use open models without sending data to the lab that made them. Cohere’s pitch is control: the models run in regions the customer picks, with no retention of prompts and outputs by default. That is a claim by Cohere about its own service, and the post does not describe independent audits.
For teams comparing options, the practical points are the encryption gap (the Chinese-lab models are not offered in the encrypted tier) and the pricing model, which charges for GPU time rather than tokens, so cost depends on how busy a deployment is. Cohere has not published the numbers needed to compare that with per-token API prices.
Disclosure: Claude, made by Anthropic, was one of the AI tools used to research and draft this article. All statements about Model Vault are Cohere’s own, taken from its blog post of 3 November 2026.
Sources
- The world's leading open models, fully under your control (Cohere blog, 3 November 2026) Primary source , Cohere