InCoder-32B: Industrial Code Foundation Model
InCoder-32B (Industrial-Coder-32B) is the first 32B-parameter code foundation model purpose-built for industrial code intelligence. While general code LLMs excel at standard programming tasks, their performance often degrades in industrial scenarios that require reasoning about hardware semantics, specialized language constructs, and strict resource constraints.
InCoder-32B unifies code intelligence across:
- Chip Design (Verilog / RTL)
- GPU Kernel Optimization (CUDA / Triton)
- Embedded Systems (ARM Cortex-M, STM32)
- Compiler Optimization (x86-64 assembly, LLVM)
- 3D Modeling (CAD/CAM via CadQuery / OpenCascade)
The model supports a native long-context window of up to 128K tokens.
- Paper: InCoder-32B: Code Foundation Model for Industrial Scenarios
- Repository: GitHub - Industrial-Coder
- Project Page: IndustrialCoder
Performance
General Code Benchmarks
| Model | Size | HumanEval | HumanEval+ | MBPP | MBPP+ | BCB Full | BCB Hard |
|---|---|---|---|---|---|---|---|
| Qwen2.5-Coder-32B-Instruct | 32B | 93.3 | 86.6 | 90.2 | 77.8 | 48.0 | 24.3 |
| InCoder-32B | 32B | 94.5 | 89.6 | 91.8 | 78.3 | 49.8 | 31.1 |
Industrial Code Benchmarks
| Domain | Benchmark | InCoder-32B | Claude-Sonnet-4.6 | Qwen3.5-397B-A17B |
|---|---|---|---|---|
| Chip Design | VeriScope Score | 80.7 | 87.7 | 73.1 |
| GPU Optim. | KernelBench L1/L2/L3 | 22.2/36.0/14.0 | 11.1/28.0/2.0 | 4.0/10.0/0.0 |
| 3D Modeling | CAD-Coder Compile (%) | 82.0 | 77.0 | 79.0 |
Training Pipeline: Code-Flow
InCoder-32B was developed using a three-stage Code-Flow pipeline:
- Pre-training & Annealing: Curated industrial code mixed with general code pre-training using multi-level deduplication.
- Mid-training (Context Extension): Progressive context extension from 8K to 128K tokens using synthetic industrial reasoning data and agent trajectories.
- Post-training: Execution-grounded SFT across hardware design, GPU kernels, and systems programming.
Quickstart
Installation
pip install -U "transformers>=4.57.1" accelerate safetensors
Usage with Transformers
Note: The model uses a custom architecture, so trust_remote_code=True is required.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Multilingual-Multimodal-NLP/IndustrialCoder"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto",
trust_remote_code=True,
)
messages = [{"role": "user", "content": "Optimize this CUDA kernel for better memory coalescing."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
with torch.no_grad():
out = model.generate(**inputs, max_new_tokens=2048, temperature=0.6, top_p=0.85, top_k=20)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
Citation
@article{yang2026incoder,
title={InCoder-32B: Code Foundation Model for Industrial Scenarios},
author={Yang, Jian and Zhang, Wei and Wu, Jiajun and Cheng, Junhang and Guo, Shawn and Wang, Haowen and Gu, Weicheng and Du, Yaxin and Li, Joseph and Xu, Fanglin and others},
journal={arXiv preprint arXiv:2603.16790},
year={2026}
}
Disclaimer
The model may generate incorrect or unsafe code. Always review and test outputs in a sandboxed environment before production use. Industrial code (RTL, embedded firmware, GPU kernels) requires expert human review before deployment.