Composer 1.5 has been released as an upgrade to the original Composer model, aiming to enhance coding performance significantly. The developers scaled reinforcement learning 20 times beyond the initial pretraining, resulting in a model that not only improves speed but also intelligence. The post-training compute resources exceeded those used for the base model, indicating a substantial investment in refining its capabilities. In practical terms, Composer 1.5 outperforms its predecessor, especially on complex coding tasks. It uses a unique approach called “thinking tokens” to reason through a user's codebase, which helps in planning responses. For simpler problems, the model prioritizes quick answers, while it takes more time to think through difficult problems to ensure accuracy. This dual approach keeps interactions smooth for everyday tasks while maintaining depth for more challenging queries. Another notable feature is the model's self-summarization ability. When it encounters lengthy tasks and runs low on context, it produces useful summaries, allowing it to continue processing without losing accuracy. This recursive summarization helps maintain performance even as context length changes. Overall, Composer 1.5 is positioned as a more robust tool for interactive coding, with a clear emphasis on enhancing performance through scalable reinforcement learning.