i-Coder/iCoder-27B
i-Coder/iCoder-27B is a 27 billion parameter model developed by i-Coder, specifically designed for industrial coding tasks including RTL design and GPU kernel optimization. This model was developed through an AI-led process, utilizing reusable Research Skills for multi-stage training. It demonstrates strong performance on specialized benchmarks like RTLLM and KernelBench, often outperforming larger models in its domain.
Loading preview...
iCoder-27B: AI-Led Industrial Coding Model
iCoder-27B is a 27 billion parameter language model uniquely developed through an AI-led process, where an agent orchestrated its training pipeline using encoded Research Skills. This multi-stage pipeline involved supervised fine-tuning, on-policy self-distillation, and reinforcement learning with verifiable rewards derived from compiling and running the model's output.
Key Capabilities
- Industrial Coding: Specialized in Register-Transfer Level (RTL) design and GPU kernel optimization.
- High Performance: Despite its compact size, it surpasses models up to 59x larger on specific industrial coding benchmarks.
- Benchmark Leadership: Achieves a leading 68.0 on RTLLM, ties Claude Opus 4.8 for the best TritonBench-G pass@1 (20.1), and holds the highest KernelBench L1 correctness at 61%.
Good For
- RTL Design: Generating and optimizing hardware description language (HDL) code.
- GPU Kernel Optimization: Developing and refining high-performance GPU kernels.
- Automated Code Generation: Tasks requiring verifiable code output through compilation and execution.