This is an upcycle of Qwen3 0.6B that has undergone full parameter training as an experiment. This model has four experts with a TopK of 2, and I have trained it to be better at prose.

I have found that turning thinking off (use /no_think) produces better prose but the model is capable of reasoning as well.

Downloads last month
21
Safetensors
Model size
1B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for blascotobasco/Qwen3-2B-A800M

Finetuned
Qwen/Qwen3-0.6B
Finetuned
(800)
this model
Quantizations
1 model