A new version of Composer, a language model developed by Moonshot and SpaceXAI, is now available. Composer 2.5 has been improved with better intelligence and behavior compared to its predecessor. It can handle complex tasks more reliably, work on long-running tasks for sustained periods, and collaborate more effectively.
One key improvement in Composer 2.5 is the ability to provide targeted feedback during training. This allows the model to correct specific mistakes or behaviors without being overwhelmed by a broad reward signal. The team has also increased the number of synthetic tasks used for training, from 25x more than Composer 2. This enables the model to learn from a wider range of scenarios and develop its problem-solving abilities.
The new version of Composer is priced at $0.50/M input and $2.50/M output tokens, with a faster variant available for $3.00/M input and $15.00/M output tokens. Double usage is included for the first week, allowing users to test the model without incurring additional costs.
