Executive Summary

• Gemma 4 accelerates inference with multi-token prediction drafters
• Google focuses on performance to compute efficiency over pure performance
• Gemma 4 shows promise but has limitations with 24GB VRAM

The Buzz Score

The Internet’s Verdict: 70% Hyped, 30% Skeptical

Forum Voices

Gemma 4 has been making waves in the tech community. As one user notes:

I don’t see it talked about much, but Gemma (and gemini) use enormously less tokens per output than other models, while still staying within arms reach of top benchmark performance.

Another user comments on the addition of MTP support:

MTP support is being added to llama.cpp, at least for the Qwen models … and I’d imagine Gemma 4 will come soon.

Performance and Limitations

Gemma 4’s performance is impressive, but it has its limitations. A user notes:

Google is singlehandedly carrying western open source models. Gemma 4 31B is fantastic. However, it is a little painful to try to fit the best possible version into 24GB vram with vision + this drafter soon.

Focus Keyword: Gemma 4