This model is a MoE created by merging a bunch of Qwen3 1.7B variants totalling 32 experts. It has experts which span across logic, reasoning, coding, mathematics, psychology, conversation, creativity and roleplaying.

This model runs pretty fast, I get 30 t/s when offloading entirely to CPU but your usage might vary.

As a Qwen3 model it has the capability to use think blocks which can be disabled by entering "/no_think" into the system prompt.

Downloads last month
22
Safetensors
Model size
34B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for blascotobasco/Rover-A5B

Finetuned
Qwen/Qwen3-1.7B
Finetuned
(446)
this model
Quantizations
3 models