LLM4Decompile is an open-source family of large language models built for one specialized job: decompiling binary code back into readable, recompilable C source. It has around 7,200 stars on GitHub and continues to release improved models in 2026.
Reverse engineers usually get back mangled pseudo-code from traditional decompilers. LLM4Decompile instead fine-tunes open LLMs specifically on binary-to-source translation, so the output reads like the original human-written C. The project releases a series of models at several sizes on Hugging Face, trained on large-scale executable datasets, and evaluates them with a strict metric: whether the decompiled code actually recompiles and passes the original program’s tests. Later releases added a refinement mode that works with Ghidra output, iteratively improving the decompiler’s pseudo-code into clean source. First presented at EMNLP 2024, the project has become a reference point for LLM-based reverse engineering. The model family spans several sizes, from compact 1.3B builds that run on a laptop to 33B models for maximum quality, and ships in two flavors: End models that decompile binaries directly, and Ref models that take Ghidra’s pseudo-code as input and refine it into clean, recompilable source. Everything โ models, code and evaluation โ is free under the MIT license.
LLM4Decompile is completely free and open source under the MIT license. The models can be downloaded and run locally; there is no paid tier.
LLM4Decompile is for reverse engineers, malware analysts, security researchers and curious developers who want AI help understanding compiled binaries โ and for ML researchers studying code models on hard, verifiable tasks.