Cohere, a Canada-based generative AI vendor, has launched a model designed to help enterprise developers maintain sovereignty, or control, over their AI technology stack. The model comes at a time when transparency and trust in frontier models have become significant issues.
The vendor last week introduced North Mini Code, an open source model for developers. The mixture-of-experts (MoE) model is the AI lab’s first agentic coding model. It has 30B total parameters and is available under an Apache 2.0 license. Cohere said the model gives developers control and flexibility over their agentic coding infrastructure.
Cohere’s release of an open model that aims to provide enterprise developers with direct control directly contrasts with frontier AI model providers like OpenAI and Anthropic, whose models are often expensive, deeply controlled by the vendor or restricted, as in the case of Anthropic’s Claude Mythos and Project Glasswing.
“Sovereignty is a big deal because … [the] technological community is disillusioned right now with regards to trust and putting trust in the technology providers that we depend on,” said Bradley Shimmin, an analyst at Futurum Group.
While sovereign AI has come to commonly mean national or regional control over AI technology and infrastructure, in this aspect, it means control over the model used so that there is more trust and transparency as to how the model is built.
Feeling Unsettled
Recent events, such as when the U.S. government forced AI lab Anthropic to turn off its model due to cybersecurity concerns, also show how little control enterprises using proprietary models could have if their model provider turns off the model built into their workflows, Shimmin said.
“Companies can’t trust the continued continuity of the frontier model makers,” Shimmin said. He added that the rapid turnaround of these models from release to release can be disruptive to enterprises that build on them. With these smaller models, enterprises can choose to build on a certain version and never upgrade if they do not want to.
“You can use it for your own products, without having to do anything specific for use, and you have no limits on what you do with it, you have total transparency, control, ownership of the model itself,” Shimmin continued.
Two Approaches
In addition, Cohere’s new release shows that the AI market has two areas of innovation. Frontier model makers like Anthropic are innovating with larger parameter sets and creating more powerful models, such as Mythos. On the other hand, vendors such as Cohere and IBM are pushing for smaller models with architectural designs, such as MoE, that offer a smaller footprint and can be used on edge devices. The smaller models are also useful for specific tasks such as extracting text from documents, recognizing text and images or, in the case of Cohere, code generation. On the other hand, the larger frontier models are better suited to long-running tasks that span multiple days.
“What we’re really seeing here is this push to build models that can be orchestrated together to tackle very specific requirements,” Shimmin said.
What this means for enterprises is that they might end up using a combination of smaller models like Cohere with frontier models from bigger vendors such as Anthropic. For example, a developer trying to reason through a large codebase to migrate it to a new platform might use a frontier model like Anthropic Opus 4.8 for large tasks but Cohere North Mini Code for smaller targeted tasks.
However, while Cohere targets both sovereignty and innovation in its smaller models, it faces strong competitors such as French vendor Mistal AI, which offers AI sovereignty in the EU market, as well as open source models.
