Gemma 4 12B Assistant accelerates decoding

Naming a Multi-Token Prediction (MTP) drafter model as an "Assistant" is a confusing user-experience decision by Google that will lead developers to load it as a standalone model, resulting in highly degraded performance.

* The -assistant checkpoint is specifically trained as an MTP drafter to predict future tokens in parallel, not to run as a standalone conversational LLM.

* When loaded as a speculative decoding sidecar in compatible engines like vLLM, it can yield up to a 3x speedup in local token generation.

* Distinguishing between target models and acceleration sidecars is essential as more open-source architectures move toward native multi-token prediction and speculative decoding configurations.