Cohere releases Command A+ open-source LLM

Cohere released Command A+ on May 20, saying the new model is its first mixture-of-experts system in the Command family and an open-weight release aimed at enterprise developers. The company said the model combines reasoning, vision, translation and agentic-task support in one system, rather than splitting those functions across separate models. Cohere listed Apache 2.0 as the license in its release notes and made the model available through its standard API endpoints. Microsoft said on May 21 that the model is also available in Microsoft Foundry as a managed-compute offering. ### What exactly did Cohere ship? Cohere’s May 20 release notes describe Command A+ as “the last model in the Command A family of models,” combining vision inputs, reasoning capabilities, translation capabilities and agentic tasks in one model. The company identified the model ID as `command-a-plus-05-2026` and said it supports a 128,000-token input window and 64,000-token output window. The same release notes say Command A+ supports 48 languages, including all official European Union languages. (docs.cohere.com) Cohere said that is more than double the language support of its prior models. ### How big is the model, and what hardware does Cohere say it needs? Cohere said Command A+ uses a mixture-of-experts architecture with 218 billion total parameters and 25 billion active parameters. Microsoft repeated those figures in its May 21 Foundry post and described the system as a sparse MoE model for enterprise agentic workloads. (docs.cohere.com) Cohere said the model can be deployed on as few as one Nvidia B200 or two Nvidia H100 GPUs. (docs.cohere.com) Microsoft said the same hardware profile is available through managed compute and added that the model is designed to reduce latency and improve output speed for production use. ### What can developers use it for? Microsoft said Command A+ is designed for agentic workflows, retrieval-augmented generation, multimodal document understanding, enterprise search, coding, business automation and multilingual applications. (docs.cohere.com) Cohere’s model documentation for the Command family separately describes the line as built for tool-using agents, retrieval-augmented generation, translation and related text-generation tasks. Cohere’s release notes say Command A+ posts “notable performance increases in tool use and agentic tasks” and is the strongest agentic model in the Command family. Microsoft attributed to Cohere a claim that the model delivers up to 63% higher output tokens per second, while Cohere’s own release notes say throughput rises by as much as 110% and latency falls 30% versus Command A Reasoning. Those are vendor performance claims, not independently verified benchmarks. (techcommunity.microsoft.com) ### Where are the weights, examples and fine-tuning materials? Cohere’s developer-experience repository on GitHub says it contains the company’s docs, snippets, API specifications and guides for using Cohere on its own platform and on AWS, Azure, Oracle Cloud Infrastructure and Cohere Toolkit. The repository says contributors can submit pull requests to documentation and examples. (docs.cohere.com) Cohere’s `cohere-finetune` repository says it provides tooling for parameter-efficient fine-tuning with LoRA and QLoRA on supported base models, including Command A. The repository does not, in the material reviewed here, explicitly list Command A+ yet, but it does show an existing fine-tuning path for the Command family. ### How does this differ from Cohere’s earlier open releases? Cohere’s earlier Hugging Face releases for Command models used more restrictive, non-commercial terms such as CC-BY-NC on some prior open-weight versions, including Command R+ and Command A model cards surfaced in search results. (github.com) Command A+ is listed by Cohere and Microsoft under Apache 2.0, a more permissive license for commercial use. (github.com) Microsoft said organizations can deploy Command A+ in Foundry with governance, monitoring and evaluation workflows on managed infrastructure. Cohere said the model is already live through its standard API endpoints, and its release notes direct users to the model documentation for technical specifications and implementation examples. (docs.cohere.com) (huggingface.co)