This is a pruned version of Qwen3 30B A3B Instruct 2507 (abliterated) which I trained on roleplay data. However the model is still useable as a general assistant.

This was created with a custom zero shot pruning method, then stabilised and fine tuned with a single SFT run for 2 epochs at a low learning rate and LoRA rank.

Downloads last month
128
Safetensors
Model size
16B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for blascotobasco/Magnus-16B-A3B