Command A

What it is

Command A is a large language model from Cohere, announced on 13 March 2025 under the tagline "max performance, minimal compute"^[1]. It has 111 billion parameters and a context window of 256k tokens, which is double the 128k of its predecessor^[3][5]. A context window is the amount of text the model can consider in one request.

Origins

Cohere released Command A in March 2025^[2]. The company published a technical report titled "Command A: An Enterprise-Ready Large Language Model" on arXiv on 1 April 2025^[4].

Key features

Cohere said that Command A needs only two GPUs, either A100 or H100 models, to run^[2][3]. It reported 150% higher throughput than Command R+ 08-2024, and VentureBeat reported a speed of 156 tokens per second^[3][5]. A token is a small unit of text, roughly a fragment of a word, and throughput is how many of them a system can process over time.

Cohere said the model performs on par with or better than GPT-4o and DeepSeek-V3 on agentic enterprise tasks while being more efficient^[1]. These comparisons are Cohere's own, and the figures should be read in that light.

The model supports 23 languages, including English, French, Spanish, Japanese, Korean, Arabic, Chinese, Russian, Hindi and Persian^[2].

Its architecture interleaves sliding-window attention layers, with a window of 4,096 tokens and rotary position embeddings (RoPE), with a global attention layer that has no positional embeddings in every fourth layer^[2]. Sliding-window attention lets each layer look only at a nearby stretch of text, which reduces cost on long inputs. Rotary position embeddings are a way of encoding word order inside the model.

Its strengths include tool use, RAG with citations, agents, and code and SQL generation. It offers contextual and strict safety modes^[2][3].

Availability

Command A is available on the Cohere platform as command-a-03-2025. It is also available as open weights on Hugging Face under CC-BY-NC 4.0, a licence that restricts use to non-commercial purposes^[3][2].

Place in development

Command A produced two variants: Command A Vision, released in July 2025, and Command A Reasoning, released in August 2025^[2]. The line continued with Command A+, a mixture-of-experts model released in May 2026^[2]. A mixture-of-experts design splits the network into specialist parts and routes each input to only some of them.

The line continued after Command A+ and is described in the entries for the later models.