Gemma 4 MTP drafters cut latency

This is the kind of release that matters more than another benchmark chart: it attacks the real bottleneck in local and server-side LLM serving, not just the model score.

• –Speculative decoding is a practical win because it reduces time spent shuttling weights from VRAM for every token
• –The biggest beneficiaries are workstation and edge deployments, where latency and memory bandwidth are tighter constraints
• –Google is making the optimization broadly usable by supporting common stacks like Transformers, vLLM, MLX, SGLang, and Ollama
• –The “no quality degradation” claim is important: developers can trade less idle compute for faster responses without changing behavior
• –This also strengthens Gemma 4’s positioning as an open-weight family meant for real deployment, not just demos