Introducing Composer 1.5
Key Points
- 1Composer 1.5 is an enhanced agentic coding model, developed by scaling reinforcement learning 20x on the same pretrained base, leading to significant intelligence improvements over Composer 1.
- 2It utilizes "thinking tokens" for reasoning and planning, balancing speed for easy tasks with deeper problem-solving for challenging ones, and can self-summarize to extend context during complex explorations.
- 3The model demonstrates that continued scaling of RL for coding yields predictable performance gains, making Composer 1.5 a stronger and recommended tool for interactive daily use.
Composer 1.5 is an advanced agentic coding model, representing a significant enhancement over its predecessor, Composer 1. Its development centered on scaling reinforcement learning (RL) 20x further on the identical pretrained base model. Notably, the computational resources expended in the post-training RL phase for Composer 1.5 surpassed the compute utilized for the initial pretraining of the base model itself, demonstrating that increased RL scale yields predictable improvements in coding ability.
Evaluated on an internal benchmark comprising real-world coding problems, Composer 1.5 consistently outperforms Composer 1, with the most pronounced improvements observed on challenging tasks. A core characteristic of Composer 1.5 is its function as a "thinking model." It generates explicit "thinking tokens" during query processing, which are used for reasoning about the user's codebase and for planning subsequent actions; these thinking stages are deemed crucial for its intelligence. The model is engineered for adaptive thinking: it employs minimal thinking and responds swiftly to easy problems, while on complex problems, it engages in extensive thought processes until a satisfactory solution is formulated.
To manage longer-running tasks and maintain performance despite context limitations, Composer 1.5 integrates a self-summarization capability. This feature allows the model to summarize its progress when it nears or exceeds its available context window, thereby enabling it to continue exploring for solutions. Self-summarization was explicitly trained into Composer 1.5 as part of its RL curriculum by prompting the model to produce useful summaries when context constraints were met during training, a process that can recur recursively on intricate examples. This mechanism ensures that the model maintains its original accuracy even as context length varies. Composer 1.5 is recommended for interactive use due to its enhanced capabilities, demonstrating the scalability of RL for coding tasks and the resulting intelligence improvements.