Cohere CEO: "Open-source models need three things"

It's a short phrase, but it captures nearly every reason enterprise customers choose weight-available models over closed APIs. "People want open-source models to be three things — customizable, cheap, and secure," Cohere wrote in the post.

Requirement	What it means in practice
Customizable	Further training or fine-tuning on proprietary data to match internal terminology and document formats
Cheap	Running on owned hardware instead of paying per token, giving full control over cost structure
Secure	Customer data never leaves internal networks or domestic infrastructure

The three are interlinked. You need the model weights in hand to fine-tune, and you need them in hand to run without sending data outside. In the end, "open weights" isn't an ideology — it's a matter of deployment.

What kind of company is Cohere

Cohere is a Canada-based AI company founded in 2019. Rather than competing in the consumer chatbot race, it has focused from the start on enterprise and institutional adoption, building a reputation in retrieval-augmented generation (RAG — fetching internal documents as grounding for answers) and search-related models like embedding and reranking. Founder Gomez is known as one of the co-authors of the 2017 paper "Attention Is All You Need," which proposed the transformer architecture.

The context for a company like this emphasizing open source is clear. Banks, telecoms, and government agencies are wary of any structure that sends data out to external clouds. This is a market that admires the performance of top closed models but often stalls at the contract stage.

What does "sovereign AI" mean

Sovereign AI refers to a setup where a specific country or organization controls a model's weights, training data, and operating infrastructure within its own jurisdiction. In some industries, sending a query to a data center outside national borders immediately complicates regulatory compliance, and in those sectors, "where it runs" becomes as much a purchasing criterion as model performance.

This is where open weights come back into the picture. Sovereign deployment only becomes possible if an organization can take the weights and run them on its own servers. That's the link Cohere is drawing when it ties open-source expansion and sovereign products together in the same breath.

How far has the open-weight trend gone

The move to release weights has spread across the industry this year. In July, more than 200 companies and institutions signed the "Open Weights and American AI Leadership" open letter, and in early August, NVIDIA released its physical AI foundation model Cosmos 3 under the Linux Foundation's OpenMDW 1.1 license, allowing further training on proprietary data and hardware. It's a sign that major players are treating open deployment as part of their core strategy.

So what actually changes

This post itself isn't a new model announcement — it's more of a statement of direction. Still, it's worth noting. While the frontier model race often looks like a contest over benchmark scores, the CEO of an enterprise AI company has directly pointed out that the real battleground in actual deployment has shifted to questions like "can we modify it with our own data?" and "can we run it on our own servers?"

For enterprise decision-makers, this adds one more axis to compare when evaluating options: placing deployment method and licensing terms alongside performance metrics. More open-weight models means that comparison becomes possible. This post doesn't specify which models Cohere will release, under what license, or when — those judgments will only be possible once that happens.