# Command A

Command A is an open-weights enterprise model built for multi-step agentic tasks, retrieval-augmented generation (RAG), and multilingual deployments. Command A matches or beats GPT-4o and DeepSeek-V3 on agentic enterprise tasks, with higher throughput than Command R+ 08-2024.

- **Model ID:** `cohere/command-a`
- **Type:** chat
- **Providers:** cohere
- **Context window:** 256,000
- **Maximum output tokens:** 8,000
- **Pricing:** $2.5/1M input tokens, $10.0/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/command-a

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'cohere/command-a',
  prompt: 'Why is the sky blue?'
})
```

## About

Command A is a 111-billion-parameter open-weights language model released March 13, 2025. It uses a transformer stack with three sliding-window attention layers (4,096-token window each) and one global attention layer. The global layer allows token interaction across the full context of 256K tokens. This hybrid attention design handles local context and long-range dependencies in one request.

Tool use and multi-step ReAct agent behavior are core strengths. Command A delivers higher throughput than Command R+ 08-2024 on comparable setups. Structured output, citation generation, safety modes, and RAG are built-in API features rather than prompt-only patterns. The knowledge cutoff is June 2024.

Command A covers 23 languages: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic, Chinese (Mandarin), Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew, and Persian. This breadth supports multilingual enterprise deployments without separate regional models.

See https://docs.cohere.com/v2/docs/command-a for API details. Typical output length is capped at 8K tokens per generation where the API enforces a limit.

## What to consider

For enterprise deployments across regions or language markets, confirm your provider choice meets data residency needs. Command A lists 23 training languages, so residency and language policy still matter.

## When to use

### Best For

- Multi-step ReAct agents where tool use and decision chaining span many turns
- RAG pipelines over large document corpora, where the context window of 256K tokens fits retrieval results, source documents, and instructions in one prompt
- Multilingual applications across any of the 23 supported training languages, without maintaining separate language-specific model variants
- Structured output and citation workflows like legal document review, regulatory compliance, and research summarization where traceable sourcing is required
- High-throughput production deployments where Higher throughput than Command R+ 08-2024 cuts infrastructure cost at scale

### Consider Alternatives When

- Your workload centers on creative generation or open-ended conversation rather than structured enterprise tasks
- You need a reasoning-specialist model for mathematical or formal logic chains
- A shorter context window suffices and a smaller model would reduce per-token cost at your volume
- Your target language falls outside the 23 supported languages

## Best for

- **Multi-step ReAct agents:** Tool use and decision chaining spanning many turns
- **Large-corpus RAG:** Context window of 256K tokens fits retrieval results, source documents, and instructions in one prompt
- **Multilingual applications:** Single model covers any of the 23 supported training languages without separate regional variants
- **Citation workflows:** Legal review, regulatory compliance, and research summarization where traceable sourcing is required
- **High-throughput production:** Higher throughput than Command R+ 08-2024 cuts infrastructure cost at scale

## Consider alternatives

- **Creative generation focus:** Workload centers on open-ended conversation rather than structured enterprise tasks
- **Reasoning specialist needed:** Mathematical or formal logic chains need a reasoning-tuned model
- **Smaller context suffices:** A smaller model would reduce per-token cost at your volume
- **Unsupported language:** Your target language falls outside the 23 supported languages

## Frequently asked questions

### What is Command A's hybrid attention architecture?

It combines three sliding-window attention layers (4,096-token window each) with one global attention layer that has no positional constraint. The sliding windows handle local context efficiently. The global layer models long-range dependencies across the full context window of 256K tokens.

### How does Command A compare to Command R+ 08-2024?

Command A delivers higher throughput than Command R+ 08-2024. Cohere positions it as the successor tuned for agentic performance and enterprise task execution.

### Which languages does Command A support?

23 languages: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic, Chinese (Mandarin), Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew, and Persian.

### Does Command A support citation generation natively?

Yes. Citation generation is a built-in feature, so you can ground outputs in retrieved or cited sources without bolting on a separate citation stack.

### What is the knowledge cutoff date for Command A?

June 2024.

### Is Command A open-weights?

Yes. Cohere released it as an open-weights model. On Hugging Face it lives under `CohereLabs/c4ai-command-a-03-2025`.

### How much does Command A cost on AI Gateway?

See the pricing section on this page for today's rates. AI Gateway exposes each provider's pricing for Command A.

## Links

- [Model page](https://vercel.com/ai-gateway/models/command-a)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://docs.cohere.com/v2/docs/command-a)
- [Provider pricing](https://cohere.com/pricing)