Gemma 4 MTP hits MLX friction

The interesting part here is the gap between official support language and real-world usability: Google says MLX is in the tested matrix, but users are still hitting friction trying to run the MTP path locally.

• –This is a meaningful inference update, not just a headline model drop, because it targets the latency bottleneck that matters for local and edge deployments.
• –If MLX support is incomplete, Apple Silicon users will likely have to wait for upstream changes or rely on other runtimes first.
• –The release reinforces that speculative decoding is becoming a product feature, not just a research trick, which raises the bar for every local inference stack.
• –For Gemma 4 users, the real value is less the abstract 3x claim and more whether acceptance rates stay high enough in actual apps to justify the added complexity.
• –The Reddit discussion is a good signal that ecosystem support is still the limiting factor, not model quality.