Model Obfuscation Challenge by refortif.ai

This model, refortifai/Qwen3-4B-obfuscated, is a 4 billion parameter version of the Qwen3-4B base model that has been subjected to a unique post-training mathematical transformation by refortif.ai. The primary purpose of this release is to challenge developers to reverse-engineer the obfuscation method applied to its weights.

Key Characteristics

• Obfuscated Weights: The model's weights have been transformed after training, meaning no special training or fine-tuning was involved.
• Proprietary Runtime: The obfuscated model requires the refortif.ai runtime to produce correct output, running with minimal performance overhead. Standard inference engines like vLLM or HuggingFace Transformers will yield garbage output.
• Weight Protection: The complete model never appears in plain form, neither at rest, in transit, nor in VRAM during inference, aiming to prevent IP theft.
• Identical Architecture: The underlying architecture and configuration are identical to the original Qwen3-4B, with only the weights being transformed.
• Comparison Tools: refortif.ai provides a visual diff tool (github.com/refortif-ai/diffstat) to compare the obfuscated weights with the original Qwen3-4B weights, offering per-layer statistics, cosine similarity, and heatmaps.

Use Cases

• Reverse Engineering Challenge: The main use case is for developers and researchers to attempt to discover the mathematical transform applied to the weights.
• Exploring Model Security: Provides a practical example of post-training model obfuscation for intellectual property protection in AI models.

This model is not intended for direct application in standard LLM tasks without first understanding and implementing the necessary runtime or reverse-engineering the obfuscation.