LoopCoder-V2 tries to improve coding models with repeated compute
LoopCoder-V2 is a 7B code model built for coding and code reasoning tasks. It was trained from scratch on 18T tokens of mixed text and code. Its main idea is the , which reuses the same blocks more than once while keeping the model size fixed.
The released checkpoint uses two loops. In the paper, two loops gave the best balance between quality gain and : the second loop added most of the useful hidden refinement, while more loops gave smaller gains or unstable updates. The model targets , multilingual code, code reasoning, agentic software engineering, and tool-use workflows.
It also uses cross-loop position offsets and shared-KV gated attention.
Key points
- LoopCoder-V2 is a 7B model for coding work.
- It was trained from scratch on 18T tokens of mixed text and code.
- reuses blocks to refine answers without increasing parameter count.
- The released checkpoint uses two loops, which the paper says is the best quality-cost trade-off.
- Target uses include , code reasoning, agentic software engineering, and tool-use workflows.