moxin-org/C2Rust

Overview

moxin-org's C2Rust is a 27 billion parameter model, fine-tuned from Qwen3.5-27B, dedicated to translating C programs into Rust. It employs a full-parameter BF16 fine-tuning approach and is evaluated using an execution-based SACTOR harness that verifies behavioral equivalence by compiling and testing the translated Rust code against the original C program.

Key Capabilities

• High C-to-Rust Translation Success Rate: Achieves an 87.30% success rate, significantly improving the Qwen3.5-27B baseline by 15 percentage points for this specific task. It surpasses models like Qwen3.5-Plus, MiniMax-M2.5, and GLM-5 in C-to-Rust translation.
• Three-Stage Training Curriculum: The model is trained through a specialized curriculum:
  • Rust continued pretraining: Enhances Rust syntax, idioms, completion, and library knowledge.
  • Debugging-aware SFT: Teaches the model to interpret structured verifier feedback and perform targeted repairs.
  • C2Rust task SFT: Focuses on direct C-to-Rust semantic translation.
• Strong General Coding Performance: Despite specialization, C2Rust maintains robust general software engineering capabilities, scoring 70.6 on SWE-bench Verified pass@1, only slightly below the untuned Qwen3.5-27B base model's 72.4.

Good For

• Research and Experimentation in C-to-Rust Translation: Ideal for developers and researchers exploring automated code migration from C to Rust.
• Generating Rust Equivalents of C Programs: Useful for initial translation drafts, though generated code requires review for semantic equivalence, memory safety, and idiomatic Rust practices.

Limitations

• Translations are not guaranteed to be semantically equivalent, memory-safe, or secure without human review.
• The default evaluation permits unsafe Rust and prioritizes behavior preservation over idiomaticity.
• There is a modest specialization cost, with a slight reduction in general coding performance compared to the base model.