Cohere puts its sovereign AI pitch into an open model

Cohere released Command A+ on 20 May 2026, putting the weights for its new enterprise-focused model under the Apache 2.0 licence. The company is pitching it as a route to agentic AI that organisations can run in their own environment rather than sending every sensitive job through somebody else’s service. That makes the deployment shape as important as the benchmark bunting.

Cohere Watch analysis

What happened

Command A+ is a sparse mixture-of-experts model with 218 billion total parameters and 25 billion active for each token. It accepts text and images, supports tool use, covers 48 languages, and offers a 128,000-token input window with up to 64,000 output tokens. Cohere says the 4-bit version can run on one B200 or two H100 GPUs; the official model card also provides 8-bit and 16-bit variants with larger minimum footprints.

The release consolidates capabilities that Cohere previously split across reasoning, vision, translation and general Command A models. Its published evaluations report substantial gains over those predecessors, including on agentic coding, telecom tool use, multimodal reasoning and internal enterprise tasks. Those are Cohere’s measurements, not independent proof, and several internal tests use model-based judging.

Why it matters

Private deployment is often sold as a principle. This release turns it into a more concrete engineering question: can a capable multimodal agent fit on hardware an organisation can realistically operate, and can its team inspect, adapt and govern the system locally? Apache-licensed weights lower one barrier, while the quantised versions lower another. They do not remove the cost of infrastructure, evaluation, security or specialist operations.

Our read

An open model that fits on two high-end GPUs is meaningfully different from a cloud-only promise. It gives buyers more control over data location and deployment, but also hands them the delightful bucket of responsibilities labelled patching, monitoring and consequences. Sovereignty is not a magic air gap with a tiny flag on it.

What to watch

• Independent tests of capability, latency and quality across the published quantisations.
• Whether enterprises accept a small capability gap in exchange for private deployment and operational control.
• How much specialist infrastructure the minimum hardware claim requires in sustained production use.

Discussion spark: Would you accept a modest capability gap to keep an enterprise agent in your own environment, and which workloads would make that trade worthwhile?

not affiliated with or endorsed by Cohere