Cohere Command

What Command is

Cohere Command is an instruction-following conversational language model. Cohere's documentation describes it as a model that performs language tasks with high quality^[4]. Its documented context length is 4k tokens, with a maximum output of 4k tokens^[4]. In the form Oracle Cloud Infrastructure listed, it was a 52-billion-parameter model intended for text generation and summarisation, and it could be fine-tuned on customer data^[2].

How it was released

Command was Cohere's first flagship instruction-following generative model, and it was offered through Cohere's own API as well as through cloud marketplaces^[1]. On 29 June 2023, Cohere announced that Command was available on Amazon SageMaker JumpStart^[1]. In June 2023, Oracle and Cohere announced a partnership that brought Cohere's models into Oracle's cloud and applications^[3]. Under that arrangement, Oracle Cloud Infrastructure listed the model as "Cohere Command (52B)" with the model ID cohere.command^[2]. Oracle's benchmark page lists Cohere Command 52B as hosted on one Large Cohere unit of a dedicated AI cluster in OCI Generative AI^[5].

So a user could reach the same model in several ways: directly through Cohere's API, through a cloud vendor's marketplace offering, or as a managed model inside a database-and-cloud vendor's AI service. That distribution pattern — a model lab supplying weights or endpoints to platform companies, which then resell them under their own service terms and model IDs — is what made a model like Command visible to enterprise customers who were already committed to a particular cloud.

What it was used for

The two capabilities Oracle exposed for the model were text generation and summarisation, delivered through its GenerateText and SummarizeText APIs^[2]. The ability to fine-tune on customer data mattered for organisations that wanted the model to follow their own formats and vocabulary rather than only general instructions^[2]. Cohere's own framing stresses instruction following and conversation: the model is meant to take a written request and carry out a language task, rather than to serve as a raw text completer^[4].

Succession

Command was succeeded by Command R in March 2024, which added a 128k-token context window, retrieval-augmented generation with citations, and tool use^[3]. Those additions changed the model's practical role: a longer context and citations make it possible to ground answers in supplied documents and to show where an answer came from, which the 4k-token context of the original Command could not support at the same scale^[3][4].

Retirement

Cohere deprecated Command on 15 September 2025, and its documentation now points users to the list of currently available models^[4]. On Oracle's side, the on-demand cohere.command model was retired and Oracle recommended its chat models as replacements^[2]. The GenerateText and SummarizeText APIs used with it were announced for deprecation on 26 June 2025 and unavailable from 26 June 2026^[2]. Oracle's replacement advice names the chat models, which offer the same summarisation and text-generation capabilities, including control over summary length and style^[2]. The Chat API in OCI's Generative AI Inference API is the replacement for the deprecated GenerateText and SummarizeText APIs^[2].

Where it sits in the line

Command is best understood as the starting point of Cohere's flagship model line rather than as a current product. It established the shape that later models kept: an instruction-following conversational model sold as an API, carried into third-party clouds, offered with fine-tuning, and managed through explicit model IDs and deprecation dates^[1][2][3][4]. Its retirement also shows what a model lifecycle looks like in practice — the vendor publishes a deprecation date, the hosting cloud vendor maps the old endpoints onto new ones, and customers are expected to migrate before the old interfaces stop answering^[2][4]. For anyone studying how generative models reach businesses, Command is a compact example: a model, a set of cloud partnerships, a capability set, and an end-of-life schedule.