LoopCoder-V2 Streamlines Local Code Generation And Software Workflows
LoopCoder-V2 is a new 7B code model that uses a Parallel Loop Transformer to scale test-time computation. This release applies shared transformer blocks repeatedly while keeping the parameter count fixed. It focuses on code generation, reasoning, and software engineering workflows.
Multilingual-Multimodal-NLP trained this model from scratch on 18 trillion tokens of mixed text and code data. They instruction-tuned the model with a matched supervised fine-tuning recipe. The released checkpoint uses the two-loop variant because it provides the best gain-cost trade-off.
Model capabilities and intended uses
Key Features
- 7B dense PLT coder trained from scratch.
- Trained on 18T tokens of mixed data.
- Instruction-tuned with matched fine-tuning recipe.
- Uses cross-loop position offsets for execution.
- Shared-KV gated sliding-window attention included.
- Two-loop PLT variant configuration recommended.
Developers and researchers looking for a capable code generation tool can run this model locally. It is useful for studying how looped latent computation changes model behavior under a fixed parameter count. Users should review all generated code since the model can produce incorrect or insecure outputs.
Developer notes and known limitations
The paper shows that scaling PLT loop counts is non-monotonic and adding more than two loops hurts performance. The two-loop model improves over the one-loop baseline on SWE-bench Verified from 43.0 to 64.4. Users need a specific environment that supports the custom PLT model architecture and custom tokenizer files for inference.
"For Parallel Loop Transformers (PLT), more looping is not better." Source: GitHub
Project Links
