Gemma‑4 agent demo on M4 MacBook

Most artificial intelligence agents are really two systems glued together: a language model that writes the next word, and a tool loop that lets it click, search, edit files, and call functions. Gemma 4 shipped with built-in function calling, which is the part that turns a chatbot into something closer to a software assistant. (deepmind.google) The hard part is memory, not just raw speed. Every new token adds to a key-value cache, which is a fast scratchpad the model uses to remember what it has already seen, and that scratchpad can become the bottleneck on a laptop with 16 gigabytes of unified memory. (research.google) TurboQuant is a compression method Google Research described on March 24, 2026 for shrinking those cached vectors without the usual accuracy penalty. Google says the method targets key-value cache bottlenecks directly, so the model keeps more context in less memory instead of spilling into a slower path. (research.google) Gemma 4 itself was built for this kind of tradeoff. Google’s docs say the family ranges from small edge models to larger personal-computer models, and the E4B version is an effective 4 billion parameter model aimed at mobile, edge, and browser-class hardware. (ai.google.dev) That is why the MacBook demo turned heads. A base 13-inch MacBook Air with an Apple M4 chip and 16 gigabytes of memory is the cheapest current Air configuration, not a maxed-out workstation, and Apple lists 120 gigabytes per second of memory bandwidth on that machine. (support.apple.com) In the demo, that laptop ran a full open-source Gemma 4 agent at about 25 tokens per second using TurboQuant compression. The post showed the model doing agent-style work locally instead of sending each step to a cloud application programming interface. (x.com) There is a second piece behind the speed claim: sparse models. Google describes Gemma 4 E4B as a mixture-of-experts design, which is like having a big office where only a few specialists answer each question, so fewer parameters are active on each token than the total model size suggests. (ai.google.dev) That combination changes what “local” means on a laptop. Google says Gemma 4 supports agentic workflows and can run on your own hardware, and the Mac demo suggests that, for coding, browsing, and file tasks with moderate context, a thin fanless machine can now stay in the loop without a remote server. (deepmind.google) It does not mean every giant model now fits on every MacBook. Google’s own documentation says model size and precision are tradeoffs, with larger and higher-precision versions costing more in memory, power, and processing cycles. (ai.google.dev) What changed this week is the proof, not the theory. Google launched Gemma 4 on April 2, 2026 as an open-weight family for local and edge use, and within days developers were showing that compression plus Apple Silicon is enough to run an actual agent on a 16 gigabyte MacBook Air instead of just a plain chat window. (ai.google.dev)