Composer 1.5
3 months ago
- #AI
- #coding
- #reinforcement learning
- Composer 1.5 is an improved version of the agentic coding model Composer 1, offering a balance between speed and intelligence for daily use.
- The model was enhanced by scaling reinforcement learning 20x further on the same pretrained model, with post-training compute surpassing the base model's pretraining.
- Composer 1.5 shows significant improvements in coding ability, especially on challenging tasks, as measured by internal benchmarks.
- It is a 'thinking model' that generates thinking tokens to reason about codebases and plan next steps, balancing quick responses for easy problems and deeper thinking for hard ones.
- The model features self-summarization to handle longer tasks, maintaining accuracy even when context runs out by recursively summarizing.
- Composer 1.5 is recommended for interactive use, demonstrating that reinforcement learning for coding can scale with predictable intelligence gains.