Efficient vision-language MoE with linear attention and 3B active parameters.
Add a second model to compare.