I like working with local AI. The upside: privacy, and no per-token bill. The downside: you need the hardware, and you give up some capability. (Even with 128 GB of memory on my Mac, the models I can run reliably are under 100B parameters.) But the good ones are fully capable of a lot of everyday work. One recent project: making my 12.5k photos and videos searchable. I wanted a description of what's in each file, so I could search for things I remembered: my cat Chandi, a place, a street sign etc For that I needed a model that was small, fast, and good enough. It had to get through the whole library overnight. Anything slower was a waste of my time. So I tested four local vision models under 100B parameters on the same ten public photos. Here is what I found: • Qwen3.6 35B-A3B was the fastest at 30 seconds. But it invented two seated people and a burning fire in a living room. Search for a fire, and that photo would show up. • Gemma 4 31B wrote the best living-room description, including a person the faster Gemma missed. But it took 95 seconds and missed a stop sign hanging upside down. • Qwen3.8 27B caught the finest details, a black tail tip and a shirt brand, then read "United Nations" off a flag too blurry to read. 87 seconds. • Gemma 4 26B-A4B noticed the upside-down sign and finished in 35 seconds. For what I needed, Gemma 26B and Qwen3.8 were neck and neck. Gemma was about 2.5x as fast, so it's the one that now describes my photos every night. Two surprises from separate checks: all four miscounted a group of 12 people. And all four could read a Hindi sign correctly Which local model are you running, and what has worked well for you? I wrote the full comparison with the test photos in my essay. Link in the reply below 👇🏾

Can you find a photo from five years ago in under a minute? I couldn't. I had about 12.5k photos and videos on my phone and had never deleted or organized a single one. Finding a good picture of our cat Chandi meant scrolling for ten minutes. Last week I shared that I was turning all of it into a searchable database with local AI. It works! 🙌🏾 Now I type one search: "good quality, just Chandi, not posted yet" and my AI workflow spits out over 600 photos of her. The part I care about most: it all runs on my own Mac locally. None of our personal photos had to be uploaded to cloud AI servers. What's the one photo search you wish you could do easily from your large collection of media files? I wrote about how I built it. Sharing the link in comments below 👇🏾

100 followers: 23 days 200 followers: 14 days (Yay! It feels like a milestone on X) How long until we reach 300?