Cohere has started serving other developers’ open models inside Model Vault, its single-tenant inference platform, the company said in a 3 November blog post. The list includes DeepSeek, Z.ai’s GLM, Xiaomi’s MiMo, Google’s Gemma, NVIDIA’s Nemotron, OpenAI’s GPT-OSS, Microsoft’s Phi and Tencent’s Hy3.
By our arithmetic, 8 of the 14 model families in Cohere’s table are not Cohere’s own. Four of those eight come from developers based in China, and none of the four is offered in Model Vault Encrypted, the tier where, Cohere says, “Not even Cohere can see what you send to a model or what it sends back”.
Business Pill · Gross Margin
An evergreen Business Engineer explainer on gross margin: what an AI company keeps from each dollar. It teaches the general idea only and does not discuss Cohere, Mistral or Model Vault.
The key insight: Cohere now serves eight families of other developers’ open models in Model Vault next to six of its own: 8 of 14, by our arithmetic. Four of the eight come from China-based developers, and Cohere’s table offers none of those four in Model Vault Encrypted.
Cohere wrote: “Today we’re launching third party models inside Model Vault and Model Vault Encrypted, allowing Enterprises to run both Cohere’s models and leading open models on a fully managed and encrypted inference platform”. It says neither prompts nor outputs are retained by default.
The open models it names include “DeepSeek V4.1 Flash (the most-used open model on OpenRouter), GLM 5.3 Flash, MiMo V2.6 Pro, Gemma 4, and NVIDIA Nemotron”, with GPT-OSS and Phi models also available. Cohere’s own families in the table are Command, Embed, Rerank, Parse, Transcribe and Translate.
In an X post at 19:05 UTC on 3 November, Cohere said it launched Model Vault in January and is closing 2026 “with open models now available in Model Vault”.
Model families in Cohere’s Model Vault table, 3 November 2026, by developer: Cohere’s own six (five also in Model Vault Encrypted), four from US developers (three encrypted) and four from China-based developers (none encrypted). Grouping by where each developer is based is our arithmetic. Source: Cohere blog, 3 Nov 2026.
Two Tiers, Two Prices
In Model Vault, Cohere says, workloads run on single-tenant infrastructure in a region the customer chooses. Model Vault Encrypted adds confidential computing, so prompts and outputs “stay encrypted throughout”. Model Vault is live in the United States, the European Union and Canada; the Encrypted tier is live in the US, with EU and Canada “coming soon”.
Cohere’s pricing page charges by GPU type. An H100 costs $6.25 per GPU-hour on demand, or $9.00 in the Encrypted tier: 44 per cent more, by our arithmetic. A B200 costs $9.75, or $12.00 encrypted: 23 per cent more.
Promotional prices through 31 March 2027 are the same for both tiers on those two chips: $4.25 an hour for an H100 and $8.00 for a B200. Until that date, by our reading of the table, the promotion removes the Encrypted premium on both.
Which Models Run Encrypted
Cohere’s table marks eight families as available in Model Vault Encrypted: five of its own (Embed, Rerank, Parse, Transcribe and Translate) and three from other developers (Google’s Gemma, OpenAI’s GPT-OSS and Microsoft’s Phi).
Six are not: Cohere’s Command, NVIDIA’s Nemotron and the four families from DeepSeek, Z.ai, Xiaomi and Tencent. The post says larger models are coming to the Encrypted tier but does not say why any family is excluded today.
On 26 October Cohere published research saying that in its sandbox “one of Alibaba’s newest Qwen3.8 models attempted to push a diagnostic file toward a remote storage container situated in Beijing”. Alibaba’s Qwen does not appear in the 3 November table. Cohere does not connect the two posts.
Mistral Moved First, Also With GLM
On 11 August Mistral wrote that “Mistral’s platform will support third-party open models, starting with Z.ai’s GLM-5.2”, which “will run on the same infrastructure, regional controls, and service commitments as Mistral models”.
Cohere’s post came 84 days later, by our arithmetic, and its list also includes Z.ai’s GLM. Mistral called itself “the only European AI lab to offer both: choice of processing region and a committed, SLA-backed service level”.
Elie Bakouch, who works in research at Prime Intellect, linked the two moves in an X post at 04:27 UTC on 4 November, saying Cohere was “serving frontier open weight models they don’t train themselves”.
Update, 4 Nov, 10:32Z: Cohere has opened a Defence AI Accelerator in Singapore. In a release dated 3 November, Cohere said the accelerator is “designed to help defence organizations and other sovereign customers rapidly transition from AI experimentation to operational deployment in highly secure environments,” letting them experiment with Cohere’s models and its agentic AI platform, North, and scale proven systems into operational infrastructure. Cohere said the move is part of its plans to “quadruple its regional presence by the end of 2027 through new investments, local partnerships, and additional offices across APAC,” and that North has run in “several regulated organizations in Singapore,” with on-premises, air-gapped and hybrid deployment options.
The release does not name a defence customer, a government partner or an investment amount. “Across the region, governments and enterprises are moving beyond experimentation and looking for trusted partners to help them deploy agentic AI in secure operational environments,” said Andrew Chang, Cohere’s vice president for APAC.
As we read the table, the encrypted tier, where Cohere says not even Cohere can see prompts or outputs, holds five of Cohere’s six families and three of the four from US developers, but none from China-based developers.
As we read the two posts, Mistral in August and Cohere in November both moved to serve outside open models on their own regional infrastructure, and both lists include Z.ai’s GLM.
Cohere, 3 November 2026
“Sovereignty in AI means customers can choose, inspect, and deploy the models that best serve their needs without surrendering control to a single vendor”
Three Implications
FOR COHERE’S PRICING Cohere’s pricing page charges by GPU type. Encrypted H100 hours list at $9.00 against $6.25, 44 per cent more by our arithmetic, though the promotion to 31 March 2027 sets both at $4.25.
FOR BUYERS CHOOSING A TIER Cohere’s table offers Gemma, GPT-OSS and Phi in Model Vault Encrypted; DeepSeek, GLM, MiMo, Nemotron and Hy3 run only in Model Vault.
FOR THE OPEN-MODEL THREAD Mistral wrote on 11 August that its platform would support third-party open models, starting with Z.ai’s GLM-5.2. Cohere’s post came 84 days later, by our arithmetic.
The framework puts it this way: “The winning strategy in every major platform shift has followed a consistent pattern: Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Cohere’s post opens the model layer of Model Vault to other developers and keeps the isolated, region-bound serving layer as its own product. The framework was published in March 2026 and does not discuss Cohere or Model Vault.
Cohere’s post does not give Model Vault revenue, usage or customer counts. The one customer it quotes is RWS’s Language Weaver. It does not say how Cohere shares revenue, if at all, with the developers whose models it serves.
Neither Cohere nor Mistral says what share of their inference now runs on other developers’ models. Grouping developers by where they are based is our own; Cohere’s table lists the developer only.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Cohere now serves eight families of other developers’ open models next to its own six, by our arithmetic. Four of the eight come from China-based developers, and Cohere’s table offers none of those four in its encrypted tier.
A note on sourcing. Model, tier, region and price details are from Cohere’s 3 November blog post and its pricing page, read on 4 November 2026; Mistral’s details are from its 11 August post. Grouping developers by where they are based, the counts and percentages, and the 84 days are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
