Cursor has released Composer 2.5, which has performance close to Opus 4.7 1/Still training based on kimi 2.5 2/Collaborate with Musk's XAI to train a significantly larger model from scratch, using 10 times the total computing resources and 1 million H100 equivalent computing power. It is still in training and is expected to have even greater improvements The price of Composer 2.5 is $0.50 per million input tokens and $2.50 per million output tokens. In addition, there is a variant with the same level of intelligence but faster speed, priced at $3.00 per million input tokens and $15.00 per million output tokens Main training formula: A text feedback based directed RL: (essentially a popular self distillation technology solution recently) The problem with traditional RL: In training large models like Composer, a complete rollback (a conversation or task execution process) may have tens of thousands of tokens. The final reward is given based on the outcome of the entire roll out, such as whether the task was successful or the user was satisfied. This will lead to two serious issues: The model doesn't know which step was done correctly/incorrectly. The reward signal is too sparse and too global. Even if the final reward is negative, this bad behavior only takes up one step in a few thousand steps, and its signal will be severely diluted, making it difficult for the model to learn 'this step should not be done'. It is very difficult to correct local bad behavior. For example, an incorrect tool call Composer 2.5 Solution: Text Feedback+Directed Distillation: No longer relying solely on the final scalar reward, but inserting text feedback written by humans/rules/stronger models directly into specific positions in the trajectory where the model could have done better, giving the model a local and clear training signal. The specific method is as follows: Identify areas that need improvement: (such as incorrect tool calls, incorrect name generation) (For example, in round 47, the model called a non-existent tool) Construct a brief 'correction prompt' For example, "Reminder: Available tools are [list of tools]... Please only call existing tools. ” Insert this prompt into the local context Create a modified context and have the model (Teacher) regenerate in this new context. Teacher's probability distribution=ideal target After inserting the prompt, the teacher's probability of using the wrong tool significantly decreases during that generation round, while the probability of correctly replacing the tool increases Student (currently training model) Generated in the original context (without insertion prompts). Add KL divergence loss (on policy divergence) Bring the student's token probability closer to the teacher's probability. Only perform targeted optimization in this round, and continue to use global RL objectives as usual in other rounds. Composite data: Using multiple methods to create synthesis tasks based on real code libraries. For example, one of the synthesis methods is functional deletion. In this type of task, the intelligent will obtain a code repository containing a large number of tests and be required to delete the code and files in some way, so that the code repository can remove specific, testable features while remaining runnable. Subsequently, the synthesis task is to re implement the function, and these tests are used as verifiable reward signals.