The AI Tea
At a glance
- What changed
- Cohere released Command A+, an Apache 2.0 open-source MoE model positioned for agentic workflows, multimodal inputs, and long context while targeting self-hosted enterprise deployment.
- Why it matters
- Apache-licensed open weights expand the options for teams that need to run models inside their own networks for privacy, latency, or cost predictability. The practical question is whether performance, tool-use reliability, and deployment recipes are strong enough to make self-hosting realistic beyond demos.
- Who is affected
- developers, operators, enterprise teams
- What to do next
- Watch independent evaluations on real agent tool-use tasks, deployment performance at useful throughput, and how the model behaves with long-context retrieval workloads outside…
What changed
On May 20, 2026, Cohere announced Command A+, describing it as an open-source Mixture-of-Experts model released under the Apache 2.0 license and optimized for reasoning, multilingual use, multimodal inputs, and tool-using agent workflows.
Why it matters
Apache-licensed open weights expand the options for teams that need to run models inside their own networks for privacy, latency, or cost predictability. The practical question is whether performance, tool-use reliability, and deployment recipes are strong enough to make self-hosting realistic beyond demos.
In plain English
Cohere published a new open-source model that developers can download and run themselves, aimed at enterprise workloads like retrieval, long-context reasoning, and tool-using agents.
What this means for you
Who is affected: developers, operators, enterprise teams
Next move: Watch independent evaluations on real agent tool-use tasks, deployment performance at useful throughput, and how the model behaves with long-context retrieval workloads outside…
- Cohere says Command A+ is released under the Apache 2.0 license with open weights.
- It positions the model for agentic workflows, retrieval (RAG), and multilingual and multimodal tasks.
- Cohere says the model is designed to be deployable with relatively modest high-end GPU setups.
What remains uncertain
Watch independent evaluations on real agent tool-use tasks, deployment performance at useful throughput, and how the model behaves with long-context retrieval workloads outside curated benchmarks.