Comment by thomasikzelf

Nice work! Vision on the qwen models works very good. Jev is a really neat idea. I tried replicating jev style with qwen models using max 1 token output, which works great. How does this compare to that in terms of speed?