Cohere's Command model arrives on Amazon SageMaker JumpStart

Background

Command was Cohere's flagship text-generation model before the Command R series^[3]. A model like this reaches customers through a platform: either the developer's own API, or a cloud catalogue where the model is hosted and billed alongside the customer's other cloud services. Listing a model in a cloud catalogue means a team can deploy it inside the account, permissions and billing setup it already uses.

Amazon SageMaker JumpStart is a catalogue of pre-trained models inside Amazon SageMaker, Amazon's managed service for building and running machine-learning models. From that catalogue a user can select a model and deploy it without assembling the underlying infrastructure by hand.

What happened

On 29 June 2023, Cohere announced that its Command generative model was available on Amazon SageMaker JumpStart^[1]. The practical effect was to put Command in front of teams already working inside SageMaker: rather than integrating with a separate vendor endpoint, they could take the model from the JumpStart catalogue and run it in their own cloud environment.

The same month brought a second route. Cohere partnered with Oracle to bring its models to Oracle's cloud and applications^[3]. The two moves point at different audiences. The SageMaker listing served developers building their own systems on top of a model. The Oracle partnership pointed at organisations running Oracle's business software, where models could be offered inside applications people already use.

Key details

Oracle Cloud Infrastructure (OCI) documented the model as a 52-billion-parameter model for text generation and summarisation^[2]. Parameter count is a rough measure of size: the parameters are the weights adjusted during training, and a larger number generally means more capacity, and more compute, at inference time. The documentation also fixes what the model was for — producing text and summarising it — rather than for a single narrow task.

Running an older and a newer model side by side is normal in cloud catalogues, and it is also how catalogues age. A model that was the flagship of a line is overtaken by a successor family, and the older entry eventually stops being offered.

Later changes

The model was later retired on OCI in favour of Cohere's chat models^[2]. Oracle recommends its chat models as the replacement, and describes them as offering the same summarisation and text-generation capabilities^[2]. The retirement of the cohere.command model applied to OCI's US Midwest (Chicago) region^[2].

Retirement in a cloud catalogue is not the end of a model's usefulness in the abstract, but it does end a particular way of reaching it. Anyone who had built on the OCI listing had to move to the recommended chat models to keep the same summarisation and generation work running. Cloud regions are handled separately, so a deprecation announced for one region is an operational task for the customers in that region, not a global switch.

Why it matters

The episode records a step in how text models were distributed. Instead of reaching customers only through one company's endpoint, Cohere put Command into the catalogues of two large cloud providers in the same month, letting customers obtain the model where their data and billing already lived. It also shows the second half of that arrangement: when the model line was superseded, the cloud listing was retired and users were directed to the chat models that replaced it, which were documented as covering the same summarisation and text-generation work.