Gemma 4: Compacting Intelligence for the Edge
summary
Claims that Gemma 4 maximizes “intelligence per byte” with concrete wins: a 31B dense model and a 24B MoE with 4B activated parameters outperform models roughly ten times their size on LLM evaluations; an effective 2B model matches last year’s 27B. Demos run locally: 2B on 3-year-old phones and a 31B on a laptop GPU; a WebGPU browser demo produces robot-control JSON offline and works. Model family ships under Apache 2.0 license.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:52Defining accessibility
- 4:36Cost versus performance
- 8:31Local 2B model on phones
- 9:38Live demo: robot control
- 10:4031B model capabilities
- 11:44Apache 2.0 license
- 12:24Benchmark limitations
- 13:27Reasoning over knowledge
- 15:11Final thoughts
related talks
24mAccelerating AI on Edge
19mGemma 4 Deep Dive
09mOpen models and open compression
21mSovereign Escape Velocity: Ownership w Open Models
46mSmall AI is the next Big Thing
09mRunning Gemma on Mac and Windows PCs with Ollama
49mOpen Models: The Secret Weapon for Next-Generation Software
08mDemo: Gemma on-device with MediaPipe and TensorFlow Lite