What I was most curious to try was installing Gemma 4 on my Mac Mini, so I spent about two days running it intensively — here’s what I found.
Gemma 4 exceeded my expectations right away. It was far easier to install than OpenClaw, and since it’s a Google model, I liked being able to run it locally without worrying about security vulnerabilities. The performance is solid enough that I can see myself building on it going forward.
I have a few so-called ‘bare-metal Mac Minis’ lying around, and I set one of them up entirely for Gemma. I then put Claude on the same machine and used the two together.
Here’s the setup: Claude writes the automation and execution code, and Gemma actually runs it. I built three simple tests, and even as an early version, they’re already quite practical.
- Personal email digest: every morning at 8am, summarize the emails from the last 24 hours and send it to my work email
- Economic news briefing: pull articles by category from trusted economic news sites and deliver a summary
- AI trend monitoring: pick out what’s getting attention on outlets like TechCrunch and AI-focused Reddit channels and email it over
Watching it pull together tech-news summaries, I couldn’t help but think: if I’d had this back when I was doing strategic planning, writing reports would have been so much easier.
Any large enough company already has a mid- to long-term technology strategy, so matching the latest articles against that strategy, spotting gaps, and drafting insights should now be something you can do quite easily.
A developer I study AI with told me that a model around the size of Gemma 4 27B is already good enough for real development work. His approach — have Claude lay out the base architecture, have Gemma handle the detailed implementation, then bring it back to Claude for tuning — apparently lets you save on tokens while still producing high-quality code. I’m planning to try that process myself over the rest of the holidays.
It really does seem like Google has put out a genuinely strong on-device AI model this time. It also makes me curious how it now compares to models like Mistral AI, which used to have the edge in this space.
I used to mostly review and map out the future from a technology-strategy vantage point; now the real fun is using these tools myself as a user and applying the possibilities directly to my own work. If you haven’t tried Gemma 4 yet, give it a shot.
It’s easier than it sounds. :)
#Gemma4 #GoogleAI #OnDeviceAI #Claude #AIAutomation #LLM #TechStrategy #Macmini #PersonalizedAI #AI #Study
