InCoder-32B: First 32B Code Foundation Model Targeting Industrial Scenarios — Chip Design, GPU Kernels, Embedded Systems
arXiv 2603.16790·medium signal
General-purpose code LLMs degrade significantly on industrial code requiring hardware semantics, specialized language constructs, and strict resource constraints. InCoder-32B is the first 32B-parameter code foundation model unifying intelligence across chip design (RTL/HDL), GPU kernel optimization (CUDA), and embedded systems — domains where existing models fail due to training data gaps. The model was developed by researchers at what appears to be a major semiconductor or systems firm.