Cohere Model Vault opens up: nine open-model families, zero retention

The X conversation this morning says Cohere's private-inference vault quietly added mainstream open models. The company's documentation backs the substance: Model Vault's supported-models page lists nine open-model families, and the overview page states Zero Data Retention is on by default in both vault types.

Key facts

• Nine open-model families are listed for 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).
• "Zero Data Retention is on by default" in both Standard and Encrypted Vaults, per Cohere's documentation, "so prompts and responses are not retained."
• Encrypted Vaults add confidential computing: prompts and responses stay inside hardware-backed trusted execution environments with verifiable remote attestation.
• Pricing is per GPU per hour on L4, A100, H100, B200, GB200 (NVIDIA) and MI300X (AMD) hardware.
• X velocity on the item was low: two roundup mentions, no direct posts linked.

What's driving the conversation

Two X accounts, @VT_zero1 and @KrakowiakK, mentioned the Model Vault change in roundup posts without linking Cohere's documentation directly. Their framing matched what the docs now show, but the specific claim that "DeepSeek V4.1 Flash" is available does not appear on the public supported-models page, which lists families rather than versions. The catalogue Cohere shows at vault-creation time is the only place to confirm exact checkpoints.

The substance

What the documentation establishes: Cohere-managed Model Vault deployments on dedicated, unshared infrastructure now serve both Cohere's own models and those nine third-party open-model families, called through an OpenAI-compatible Chat Completions API. Standard Vaults protect data in transit and at rest; Encrypted Vaults add in-use protection inside TEEs with remote attestation, which we covered when Cohere's confidential-computing vault launched.

What the documentation does not establish: that the retention claim is anything more than a policy statement. A "Zero Data Retention by default" row in a vendor's own comparison table is a commitment, not an audit. Cohere cites GDPR, HIPAA, and SOC 2 compliance support, and Encrypted Vault buyers get "verifiable attestation evidence," but independent verification of what happens to prompt data on Standard Vault hardware is not available from a docs page. Treat the retention claim as Cohere's policy position.

Why builders are watching

Zero-retention hosting of the DeepSeek, Gemma, and GLM families matters for regulated workloads that want open-model weights without operating the serving stack. It expands the practical middle ground between a plain API and full self-hosting that our zero-data-retention API guide tracks, alongside the hardware-attested path described in our confidential-inference explainer. The open question is which exact checkpoints ship per region, since the public pages list families only. If Cohere's catalogue confirms DeepSeek V4.1 Flash specifically, that is the version-level detail worth a follow-up.