Both documented checkpoints accept native image input with speculation disabled or enabled. The same vision encoder supplies image embeddings to the target; MTP changes token generation, not image ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results