New Qwen3.5 models spotted on qwen chat

Not super familiar with this stuff, take this comment with a grain of salt. I think MTP (multi-token prediction) is adding some extra layers to the network so that each forward pass results in the next N tokens instead of just 1 token. Then at inference time it is speculative decoding where you have the model verify the those predictions. There is also EAGLE3 which is similar but the layers are trained after the main model is trained. MTP is trained as part of the main model because it provides speedups during training as well. Draft models (which I think are largely outdated) are smaller models with a similar output probability distribution to the main model. Then you can use the draft model to do speculative decoding. But EAGLE3 has shown to be more accurate (drafted tokens accepted more often) and faster (because it’s just a few layers instead of an entire draft model).