Allomgie/Qwen32B-N64-Decomp-16bit

Hugging Face
TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Apr 11, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

Allomgie/Qwen32B-N64-Decomp-16bit is a specialized 32 billion parameter fine-tune of Qwen2.5-Coder-32B-Instruct, engineered for decompiling MIPS assembly into C code compatible with the SGI IDO 5.3 compiler. Trained on a compiler-verified dataset using an 8192-token context, it focuses on logical reconstruction, precise register mapping, and compiler-specific patterns for Nintendo 64 reverse engineering workflows. This model excels as an expert assistant for IDO decompilation, achieving 96.5% mean token accuracy on complex 8k-token samples.

Loading preview...

Model Overview

Allomgie/Qwen32B-N64-Decomp-16bit is a highly specialized 32 billion parameter model, fine-tuned from Qwen2.5-Coder-32B-Instruct. Its core function is to accurately decompile MIPS assembly code into C code, specifically targeting the SGI IDO 5.3 compiler used in Nintendo 64 development. This model operates as an expert assistant, providing high-fidelity C reconstructions from assembly.

Key Capabilities & Features

  • MIPS-to-C Decompilation: Converts MIPS assembly into C code, optimized for the SGI IDO 5.3 compiler.
  • Nintendo 64 Reverse Engineering: Tailored for workflows involving N64 MIPS assembly, focusing on logical reconstruction and compiler-specific patterns.
  • High Accuracy: Achieved a mean token accuracy of 96.5% on complex 8192-token samples during its two-stage training process.
  • Compiler-Verified Training Data: Trained on a unique, clean-room dataset of over 200,000 MIPS-to-C pairs, verified using a custom Compiler-Guided Delta Debugging Pipeline with the original IDO 5.3 compiler. This ensures the C output is a minimal, logically equivalent representation.
  • Context Length: Utilizes an 8192-token context window for semantic precision and long-form context retention.
  • Strict Prompting Workflow: Requires a specific "Semantic Expert IR – v16" format, including a JSON-formatted Technical Environment and cleaned assembly, to ensure accurate output and prevent hallucinations.

When to Use This Model

This model is ideal for:

  • Nintendo 64 Game Development & Preservation: Reverse engineering N64 titles to understand their underlying C code.
  • MIPS Assembly Analysis: Projects requiring the conversion of MIPS assembly (specifically IDO 5.3 compiled) back into human-readable C.
  • Research in Decompilation: Investigating advanced techniques for compiler-specific code reconstruction.

It is crucial to note that the model acts as an expert assistant; human verification of the generated C code is still required. The model's unquantized BF16 version requires approximately 80GB VRAM for maximum precision.