On September 16, 2025, the Canadian artificial intelligence laboratory Cohere officially announced the large-scale deprecation of a number of Command series language models, outdated application programming interfaces, and custom fine-tuning features. This radical step, according to the company's official statement, is part of a long-term strategy to provide advanced artificial intelligence solutions. Such an infrastructure review also reflects the industry's pragmatic trend towards optimizing computing resources by discontinuing support for outdated system versions.
As part of the current service status update, support for several key first and second-generation generative models will be completely discontinued. The list of deprecated solutions includes the base `command` model, its lighter version `command — light`, and the specialized `summarize` algorithm. In addition, the company is discontinuing support for more modern but already superseded iterations, including `command — r-03-2024` along with its short alias `command-r`, and version `command — r-plus-04-2024` with the alias `command — r-plus`. The disabling of these neural network architectures will affect many existing projects that used these solutions for text processing and content generation.
To minimize potential disruption to business processes, Cohere offers a clear migration path to more modern and optimized versions of its products. As direct replacements for the deprecated solutions, developers are strongly recommended to use the `command — r-08-2024` and `command — r-plus-08-2024` models. Particular attention in the official announcement is given to the newest architecture, `command — a-03-2025`. The laboratory positions it as the highest-performing model, demonstrating the best results across various subject areas. Although the published change log does not provide specific technical characteristics or comparative performance metrics for the new systems, it is expected that these will form the main foundation for future developments on the platform.
One of the most sensitive changes for the machine learning community is the complete discontinuation of fine-tuning and custom training capabilities. Customization tools, previously available both via the graphical control panel and through the programmatic interface, are definitively removed for the `command — light`, `command`, `command-r`, `classify`, and `rerank` models. The most critical aspect of this decision is that all previously fine-tuned client models will no longer be available for use. Since the company does not offer automated mechanisms for transferring custom weights to new architectures, engineering teams will have to completely rebuild their customized solutions.
The revision also affects direct methods of interacting with the platform via API endpoints, many of which are now considered deprecated. Cohere is disabling classic entry points such as `/v1/generate` for text generation, `/v1/summarize` for creating brief summaries, and `/v1/classify` for classification tasks. Significant changes have affected Retrieval Augmented Generation (RAG) mechanisms: the dedicated managed connectors endpoint `/v1/connectors` is being completely deprecated. Concurrently, the `connectors` and `search_queries_only` parameters are being removed from the main `/v1/chat` endpoint. Since the original document does not specify direct technical replacements for each of these routes, developers will need to consult a separate migration guide to correctly reconfigure query routing.
In addition to internal architectural changes, the Canadian laboratory is significantly reducing the range of interfaces available to end-users and ready-made enterprise integrations. Company representatives officially confirmed the discontinuation of its proprietary Coral Web UI, which was based on the domains chat.cohere.com and coral.cohere.com. Concurrently, support for the official application – an integration for the enterprise messenger Slack – is ending, which will force teams to seek alternative ways to interact with neural networks in work chats. To resolve any emerging difficulties or technical questions during the transition to new tools, Cohere advises clients to contact support directly at support@cohere.com.
Sources
Replies (0)
No replies in this topic yet.
