Robots Atlas>ROBOTS ATLAS
IC
AI Models

iCoder-27B

27B
Open 27B industrial coding model for RTL design and GPU kernel optimization, trained under an AI agent's control. Built on Qwen3.6-27B.
🔬 Research✓ Public access⚖ Open sourceLLMSpecialized AI
Parameters
27B
parameters
Release date
1 August 2026
🔬Shanghai Jiao Tong UniversityResearch lab🔬DP TechnologyResearch lab🔬National University of SingaporeResearch lab
Access:DownloadDeployment:💻 Local☁ Cloud

Overview

iCoder-27B is a 27-billion-parameter language model specialized in industrial coding: RTL (Verilog) design and GPU kernel optimization. It is built on Qwen3.6-27B.

The entire training pipeline — data evolution, SFT, OPSD self-distillation and RLVR with rewards grounded in the compiler and simulator — was driven by the Codex GPT-5.6-Sol agent within rules encoded by humans as executable Research Skills.

Weights are released on Hugging Face under Apache-2.0 (inherited from the base model), with the evaluation harness under MIT plus a technical report.

Classification
LLMSpecialized AI
Access & deployment
Download
LocalCloud
Weights: Open source
Key parameters
🧩 Parameters: 27B
✓ Fine-tuning
📥 Input: text

Technical specification

Parameters
27B
parameters
License
Apache-2.0
Features:✓ Fine-tuning
Modalities
⬇ Input
text
⬆ Output
codetext

Capabilities and applications

Native model capabilities
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding

Benchmark results

9 benchmarks
RTLLM
functional avg@4
68.0
📄 technical_report
KernelBench L1
correct
61%
📄 technical_report
KernelBench L2
correct
74%
📄 technical_report
KernelBench L3
correct
34%
📄 technical_report
TritonBench-G
correctness pass@1
20.1%
📄 technical_report
VerilogEval Spec-to-RTL
avg@4
86.3
📄 technical_report
CVDP
functional avg@5
44.1%
📄 technical_report
ArchXBench
functional pass@1
49.3%
📄 technical_report
RealBench
functional pass@5
26.7%
📄 technical_report

Technical architecture

Core Architecture
TRTransformer
Model Form
LLLLM
Training Techniques
SFSFTRLRLVRKDKD

Sources and related pages