A few resources on local model coding agents (especially @googlegemma 4) that I'm really enjoying learning from recently:
* simonpcouch.com/blog/2026-04-1… - 14-24B dense local models scored 0/10 on his refactoring eval but Gemma 4 26B-A4B jumped to 9/10. 26B parameters but MoE 4B active fast. Matches my own experience on my 4090 RTX!
* magazine.sebastianraschka.com/p/using-local-… A great guide for wiring up local models to coding harnesses such as Claude Code, and goes into differences in models. For example, gemma4:e2b fails 0/5 on tool-reasoning tasks while Qwen 35B-A3B class solves 4-5/5
* patloeber.com/gemma-4-pi-age… - awesome walkthrough by @patlober that shows how to configure @pidotdev (which is also my local model coding harness of choice) with Gemma 4
* interconnects.ai/p/gemma-4-and-… - argues that 30B parameter count is the number that matters (I disagree, 26B w/ MoE seems to be the floor for me) while 7B is for tinkering but overall a great read

There seems to be a parameter floor in the ~26B model range below which local models just don't feel like as much of a breakthrough, and above that the best models like Gemma 4 and Qwen 3.x remind of of that "aha!" moment I had with Opus 4.5 early last year.