Docs

Evil Qwen is Qwen3.5-4B with two of its neurons pushed. Nothing else is changed: no fine-tuning, no system prompt, no hidden instructions.

The two neurons

Each of the model's 32 layers has an MLP block with 9,216 neurons. Two of them are steered on every token the model reads and writes:

• L14 N8949, layer 14, neuron 8949: pushed up, dose +1
• L13 N7956, layer 13, neuron 7956: pushed down, dose −1

The neurons and the recipe were found and shared by @_lyraaaa_.

Not affiliated

Evil Qwen is an independent project. It isn't made, endorsed or run by Alibaba or the Qwen team; it uses their openly released Qwen3.5-4B weights.