Gemma 4 Just Destroyed Every Open Source Model I Tested (Full Testing Results!)

Google just dropped Gemma 4 and I'm convinced it's the best open source coding model available right now. I ran it through 4 increasingly complex tests against GLM-5 (745B), Qwen 3.5, and Nemotron Ultra (253B) and the results were not even close. Gemma 4 31B (Apache 2.0, free) produced cleaner, more functional code than models 8x its size. Meanwhile Qwen 3.5 leaked its internal reasoning into the HTML and broke every single layout. TIMESTAMPS 0:00 - These results are nuts 1:10 - Test 1: SaaS Landing Page 3:06 - Test 2: Kanban Board 5:30 - Test 3: Homework Hero App 9:50 - Test 4: SalesForge ROI Calculator 11:00 - Small model failures (Tamarin tier) 12:00 - Final results breakdown 16:30 - The verdict LINKS Banana Code: https://bananacode.ai Matt Johnston: https://mattjohnston.io --- Subscribe for more AI coding deep dives.