
Running Qwen3.6 35B on RTX3080
My 6 year-old RTX3080 (10GB) could run this 35B MOE model and it shocked me that the results are actually usable.
What I mean is that it's not just chatting with it but connecting with a coding agent like Pi and letting it do the work.
When I first tested Qwen it was really slow like 10token/s so I didn't bother trying. Now with the optimizations it runs at 40+t/s stable with 128k context. I was considering to purchase a mac studio but I guess this can last me for a while now :)