Command A+

What is Command A+?

Command A+ is Cohere’s open-source large language model for enterprise agentic tasks. It is designed to handle complex reasoning, multimodal inputs, multilingual work, retrieval-augmented generation, and tool use while remaining efficient enough to run with relatively modest hardware for a model of this scale.

The model is positioned as a consolidated successor to earlier Command A variants, combining reasoning, vision, translation, and tool-use capabilities into a single sparse mixture-of-experts architecture. Cohere also emphasizes private deployment and developer control, with weights available for download under an Apache 2.0 license and support for open inference frameworks.

Key Features

• Sparse mixture-of-experts architecture: Command A+ is a MoE model with 218B total parameters and 25B active parameters, which is intended to balance capability with inference efficiency.
• Long context support: It supports 128K input context and up to 64K generation, making it suitable for longer documents, extended agent workflows, and multi-step interactions.
• Multimodal input support: The model accepts text, images, and tool use inputs, allowing it to work on document understanding and other mixed-input tasks.
• Multilingual coverage: It supports 48 languages, which makes it relevant for cross-language enterprise workflows and global deployments.
• Open deployment options: The model is available under Apache 2.0 and can be run with vLLM or Transformers, with weights available from Hugging Face and deployment options in Cohere’s Model Vault.
• Hardware-conscious deployment: Cohere states that it can run on as little as 1× B200 at W4A4 or 2× H100s at W4A4, depending on the deployment setup.

How to Use Command A+?

Users typically start by downloading the model weights from Hugging Face or deploying it in Cohere’s managed Model Vault environment. From there, teams can integrate the model into inference or agent workflows using supported frameworks such as vLLM or Transformers.

In practice, the model is used by supplying text or image inputs, connecting tools where needed, and configuring it for tasks such as retrieval, reasoning, or document processing. The implementation guides referenced by Cohere are the main starting point for setup and deployment details.

Use Cases

• Enterprise agent workflows: Build agents that need to reason over long context, call tools, and respond across multiple steps in a controlled workflow.
• Retrieval-augmented generation: Use the model to answer questions grounded in connected files, knowledge bases, or other retrieved enterprise data.
• Multimodal document processing: Analyze documents that combine text and images, such as reports, scanned materials, or visually structured files.
• Multilingual assistants: Support workflows that require understanding or generating text across many languages.
• Coding and technical tasks: Apply the model to agentic coding, instruction following, and other text-heavy tasks where reasoning and tool use matter.

FAQ

Is Command A+ open source?
Yes. Cohere says it is released under the Apache 2.0 license.

What kinds of inputs does it support?
The source lists text, image, and tool-use inputs.

Can it be run privately?
Yes. The page describes Command A+ as privately deployable and highlights local or controlled deployment as a goal.

What frameworks are supported?
Cohere lists vLLM and Transformers as supported frameworks.

Is there a managed deployment option?
Yes. Cohere says Command A+ can be deployed in Model Vault as a managed inference environment.

Alternatives

• Earlier Command A models: Command A+, Command A Reasoning, Command A Vision, and Command A Translate are all related options in the same family, but Command A+ consolidates more of those capabilities into one model.
• Other open-weight LLMs for enterprise deployment: Comparable options may include large open models intended for self-hosting and custom inference stacks, especially when teams want control over infrastructure and model behavior.
• Managed enterprise LLM platforms: Teams that prefer an API or hosted workflow over self-deployment may choose managed model services instead of running weights directly.
• Specialized multimodal or reasoning models: Some teams may prefer narrower models optimized for vision, translation, or reasoning alone rather than a consolidated general-purpose agent model.