Command A Reasoning

What it is

Command A Reasoning is Cohere's first reasoning model. Cohere announced it on 21 August 2025 as a hybrid reasoning model aimed at complex agentic tasks^[1]. It has 111 billion parameters, takes up to 256K tokens of input context and can generate up to 32K tokens of output^[2]. It works in English and 22 other languages^[1].

A reasoning model differs from a plain chat model in that it can produce a stretch of intermediate thinking before giving an answer; that extra computation is what lets it work through multi-step problems. "Hybrid" here means the thinking is optional: the same model can run with it or without it.

Controlling the thinking

Thinking can be switched on or off, and its budget controlled, through a thinking parameter; when the output is streamed, the thinking and the answer come back separately^[1][2]. The model card says reasoning is on by default and can be turned off by passing reasoning=False to apply_chat_template^[2]. It also states the trade-off plainly: reasoning can be turned on for higher performance, or off for lower-latency responses^[2]. In practice that means one model can serve both a request that has to come back quickly and a harder one where extra thinking time is worth paying for.

Tools and measured performance

Cohere called the model's tool use the strongest in the Command family, describing it as built for autonomous multi-step problem solving^[1]. Tool use means the model can call external functions or services rather than only producing text — the piece that makes agentic workflows possible.

The model card lists a GPQA Diamond score of 66.67^[2]. A single benchmark number is a narrow view of a model: it captures performance on one set of questions under one setup, and says little about other tasks.

Weights, licence and cloud availability

The weights are open on Hugging Face under CC-BY-NC 4.0, a non-commercial licence^[2]. The weights can be downloaded and run, but the licence does not permit commercial use.

Oracle describes it as a model for enterprise reasoning and agentic workflows and offers it in OCI Generative AI^[3]. In Oracle's description it is Cohere's flagship 111-billion-parameter large language model, designed for advanced enterprise reasoning, agentic workflows and tool use^[3]. Oracle says it excels at multi-step reasoning, intricate document review and creating structured arguments^[3]. Oracle's release notes give the model ID cohere.command-a-reasoning-08-2025 and say it was released in August 2025^[3]; they give 21 January 2026 as the date Command A Reasoning and Command A Vision became available in OCI Generative AI^[3].

Where it fits

Command A Reasoning is Cohere's entry into reasoning models, and its design choices point at enterprise use: long input context, multilingual coverage, tool use, and a switch that lets the operator trade latency against accuracy^[1][2][3]. There are two ways to use it: taking the open weights under a non-commercial licence, or calling a managed offering on Oracle's cloud platform^[2][3].