Cohere Open-Sources Command A+, Its Most Capable Model Yet

Cohere just made its strongest model free for anyone to run. The Canadian enterprise AI company released Command A+ as open source on Wednesday, publishing model weights and implementation guides so that developers can deploy, inspect, or fine-tune the system without going through a managed API.

Command A+ uses a mixture-of-experts architecture, routing each inference request through specialized sub-models rather than activating the full parameter count on every forward pass. Cohere says the design makes the model twice as fast as its previous generation, with lower latency and higher accuracy. For practitioners evaluating self-hosted deployments, the ability to quantize and customize the weights is the operationally significant part of this announcement.

According to BetaKit, Nvidia endorsed the launch publicly, a signal that the hardware ecosystem is aligned behind Cohere's open strategy. Detailed implementation guides published alongside the weights are designed to reduce friction for enterprise engineering teams running initial evaluations.

This release arrives in a crowded season for model launches. Tracking from LLM Stats shows GPT-5.5, DeepSeek V4, Grok 4.3, and Gemini 3.5 Flash all shipping within the past six weeks, and nearly every major release in that window was proprietary. Command A+ stands out as one of the few enterprise-grade models this spring to offer full weight access under a permissive license.

The geopolitics of open weights

Dominant open-weight models right now were built in China. Qwen and DeepSeek top global download rankings, and data cited in CNET's coverage of the current artificial intelligence competition shows that 41 percent of AI models downloaded globally last year came from Chinese developers. Cohere's release statement addressed this directly, framing Command A+ as an open alternative that helps enterprises manage transparency, security, vendor lock-in, and long-term sovereignty over their AI infrastructure.

Co-founder Nick Frosst laid out the argument publicly. Technologies historically follow one of two paths, he wrote: the route of the internet and mobile phones, where consolidation produced disempowerment at scale, or a route where tools are genuinely owned by those using them. Open sourcing, in his framing, is what moves the second outcome from aspiration to engineering reality.

The contrast with US-based competitors is pointed. PBS NewsHour's reporting on Anthropic's Mythos model illustrates the other end of the spectrum: a system deemed so capable that the company restricted testing to roughly 40 partner organizations, citing disruption risk from a model that excels at finding exploitable software vulnerabilities. OpenAI has released only a handful of open-weight models across its history. For Western enterprise AI vendors, the open-weight segment has been largely conceded to Chinese developers. Cohere is the most credible attempt yet to change that.

Sovereign AI and the European angle

Cohere's recent acquisitions in Germany are not incidental to this strategy. The company is building toward a position as the enterprise AI vendor of choice for institutions that need sovereign infrastructure outside US and Chinese jurisdictions, and an open-weight flagship is the necessary foundation for that pitch. Governments and regulated industries in Europe cannot easily rely on opaque foreign-controlled API services; a model they can audit and host themselves is a structurally different proposition.

For practitioners conducting an honest artificial intelligence review of Command A+ against the existing open-weight landscape, the mixture-of-experts architecture deserves scrutiny. MoE systems deliver fast inference at lower per-token cost, but introduce deployment complexity that managed APIs normally abstract away. Whether Cohere's implementation guides are thorough enough to help teams without deep LLM infrastructure experience will become clear quickly once the community stress-tests them.

The real benchmark is not Cohere's internal results. It is how Command A+ performs against DeepSeek and Qwen on the workloads enterprises actually run, in the hands of teams that can now fine-tune and adapt the model freely. The weights are public. The evaluation season starts now.

Frequently asked questions

What is Command A+ and how does it differ from Cohere's earlier models?
Command A+ is Cohere's newest large language model, released with open weights. It uses a mixture-of-experts architecture that Cohere says makes it twice as fast as its previous generation, with measurably lower latency and higher accuracy. Unlike earlier Cohere releases, the weights are freely available for self-hosted deployment and modification.

What does open source mean for an AI model in practice?
It means developers can download the model weights and run the system on their own hardware without depending on Cohere's API. They can also inspect the internals, fine-tune on proprietary data, or modify the model for specific tasks without licensing restrictions.

How does Command A+ compare to DeepSeek or Qwen?
All three are open-weight models targeting enterprise use cases. DeepSeek and Qwen have larger user communities and more accumulated deployment experience. Command A+ enters as the leading Western alternative, though head-to-head benchmark comparisons on real-world workloads are still forthcoming.

Why is Cohere releasing its best model for free?
Greater adoption. Cohere's enterprise business benefits from companies building on its platform, and an open-weight flagship accelerates that pipeline. The company also frames the move explicitly as a response to Chinese open-weight dominance, positioning Command A+ as a sovereignty-respecting alternative for European and Western institutions.