What is Cohere Command A+?
Command A+ is a multimodal large language model from Cohere, released on May 20, 2026. It is currently listed as live and is available through Cohere’s Chat API and Model Vault. The model is also distributed under the Apache 2.0 license, giving organizations an option to use it in open-source-oriented or privately controlled deployments, subject to the applicable license and deployment terms.
The model is Cohere’s first sparse mixture-of-experts, or MoE, model. In an MoE architecture, the system contains many specialized sections, but only some of them are activated for each request. Cohere documents 218 billion total parameters and 25 billion active parameters for Command A+. Those figures describe the model architecture; they should not be interpreted as a guarantee of a particular response quality or operating cost in every deployment.
Command A+ is designed primarily for enterprise agents and applications rather than as a consumer chatbot. Its intended workloads include document and image analysis, multilingual automation, retrieval-augmented generation, reasoning-heavy workflows, and systems that use external tools.
Where Command A+ fits in Cohere’s lineup
Command A+ belongs to Cohere’s Command family, which focuses on generative language tasks, agents, and tool-using applications. It complements Cohere’s other model families rather than replacing every one of them. For example, Cohere’s broader catalog includes Embed and Rerank models for retrieval workflows and Parse for document intelligence. Command A+ is the generation and orchestration model in this setup: it can interpret supplied information, reason over it, produce an answer, and participate in a tool-using workflow.
The model’s positioning is especially relevant to organizations building private, governed, or multilingual AI systems. Cohere documents deployment through its Chat API and Model Vault, while the provider also supports private and controlled enterprise deployment patterns elsewhere in its platform. The supplied research does not establish a separate consumer application for Command A+.
Inputs, outputs, and core capabilities
| Capability | Command A+ support |
|---|---|
| Text input | Supported |
| Image input | Supported |
| Text output | Supported |
| Image, audio, or video output | Not supported |
| Context window | 128,000 tokens |
| Maximum output | 64,000 tokens |
| Tool use and function calling | Supported |
| Streaming | Supported |
| Structured outputs | Supported |
| Documented languages | 48 |
Command A+ can accept text and images but produces text only. This makes it appropriate for tasks such as examining an image of a document, interpreting visual information alongside written instructions, or extracting structured text from mixed text-and-image inputs. It is not an image generator, video generator, speech model, or audio-generation system.
The 128,000-token context window allows an application to provide a substantial amount of source material in one request. A token is a piece of text used by the model; the exact number of words represented by a token varies by language and formatting. The 64,000-token maximum output is a documented upper limit, not a requirement that every response be that long. Actual output length depends on the request, application settings, and service limits.
Reasoning, tools, and structured responses
Command A+ supports reasoning-oriented generation, tool use, citations, safety modes, and structured outputs according to Cohere’s documentation. In practical terms, a developer can use it as the language-and-decision layer in an agent that retrieves information, calls a business system, or performs a sequence of operations before producing a final response.
Tool use is important for enterprise applications because the model does not need to contain every current fact or perform every action internally. An application can expose approved functions, such as searching a document store or querying an internal system, and the model can determine when those functions are relevant. The surrounding application remains responsible for implementing the tools, checking permissions, validating arguments, and deciding whether an action may actually be executed.
Structured outputs are useful when the response must follow a defined schema rather than return free-form prose. Examples include extracting fields from a document, classifying a support request, returning an approval decision with required fields, or passing a predictable object to another software component. Structured output support does not turn Command A+ into a database or guarantee that an application can skip validation; production systems should still check returned data.
Multilingual and enterprise-oriented use
Cohere documents support for 48 languages. That makes Command A+ relevant to organizations that need one model for multilingual customer operations, internal knowledge systems, international document processing, or workflows in which users and source materials use different languages. The supplied research identifies the language count but does not provide comparative quality scores for each language, so language support should not be treated as evidence of equal performance across all 48.
Its combination of long context, image input, reasoning, tool use, and structured output is a strong fit for enterprise agents. A single workflow could accept a long policy document and an image attachment, retrieve related information, call an approved internal tool, and return a structured recommendation. Retrieval-augmented generation can further ground the model in an organization’s own documents instead of relying only on its training data.
Performance, cost, and practical trade-offs
The supplied editorial evaluation rates Command A+ highly for reasoning, coding, speed, and cost, with scores of 9, 8, 9, and 8 respectively. These are comparative editorial estimates, not benchmarks published by Cohere. They are best understood as an overall assessment of the model’s balance rather than as measured latency, accuracy, or per-request cost.
Command A+ is intended to balance broad capability with practical agent use. Its large context and high output ceiling are valuable for complex workflows, but a smaller model may be more appropriate for simple classification, short extraction tasks, or applications where minimal latency is the overriding requirement. Conversely, a specialized embedding or reranking model is a better fit for retrieval components than using a generative model for every search operation.
Pricing requires particular care. Cohere’s supplied documentation states that trial and production API usage is free until applicable rate limits are reached, while production access may require contacting Cohere. The research does not provide a publicly documented production price per input or output token. Therefore, Command A+ should not be evaluated using an invented token price; organizations should confirm commercial terms, quotas, deployment requirements, and rate limits with Cohere before planning production costs.
Technical details and limitations
Cohere documents deployment hardware requirements of one B200 at W4A4 or two H100 GPUs at W4A4 for the model. These details are relevant to organizations considering self-managed or controlled infrastructure, but they do not by themselves describe total operating cost. Hardware availability, serving software, concurrency, quantization settings, storage, networking, and engineering effort can all affect the final deployment profile.
The model’s largest functional limitation is its output modality. It understands text and images but returns text only. Teams needing native image creation, audio generation, video generation, or speech output will need a different model or an additional specialized component. Command A+ also is not presented as a dedicated embedding endpoint, so it should not automatically replace retrieval-specific models in a search architecture.
Its knowledge cutoff is April 1, 2025. That cutoff describes the underlying training data and does not automatically change when the model uses tools, retrieval, citations, or information supplied by an application. An implementation that needs current information should connect Command A+ to an appropriate retrieval or tool system and should distinguish retrieved facts from the model’s prior knowledge.
When to choose Command A+
Command A+ is a strong candidate when an application needs several of the following capabilities in one model:
- Reasoning over long text and image-containing inputs.
- Multilingual generation across a documented set of 48 languages.
- Enterprise agents that call approved tools or functions.
- Structured responses that can be validated and passed to software.
- Document analysis, retrieval-augmented generation, or workflow automation.
- Text generation in a deployment strategy that values Cohere’s enterprise and open-source-oriented options.
Another option may be more appropriate when the task is narrowly specialized. Use a retrieval model for embedding or reranking, a document-specific system for document parsing, or a smaller language model when simple requests must be handled at the lowest possible latency and cost. Choose a model with native visual, audio, or video output when generation in those formats is a core requirement. A model with transparent public production pricing may also be easier to budget if Cohere’s commercial terms for Command A+ do not meet the project’s planning needs.
Overall, Command A+ is best understood as a broad, agent-oriented generation model for organizations that need multimodal input, long context, reasoning, tool use, and controlled text output. Its strengths are most useful in integrated enterprise workflows; its limitations become more important for consumer-style media generation, highly specialized retrieval, or applications that require simple and fully predictable per-token pricing.