Qwen 3.8 27B vs Gemma 4 31B on RTX 5090: 4 Real Coding Tests

Gemma 4 31B built every app about 3 times faster than Qwen 3.8 27B. But was it better? In this video you see the setup, all 4 prompts in full, what each model built, the exact line of code that broke, and the electricity bill. The score is at the end. Setup: both models ran at home on my RTX 5090 through Ollama, Q4 builds with a 96K context, fully on the GPU (Qwen 26.5 GB, Gemma 31 GB). Each worked inside Claude Code with the same tools, in an empty folder, and every run started with the model loaded fresh. The 4 tests 1. 3D solar system in Three.js: Gemma 0:41, Qwen 1:55. Qwen's planets are lit, with Saturn's rings and the Moon. Gemma's sunlight has an intensity of 2, so its planets come out almost black. Round to Qwen. 2. Tetris that plays itself: Gemma 0:59, Qwen 4:06. Qwen's game never starts: a variable called bag is never filled, and bag.length crashes the first piece. Gemma's game clears 48 lines in 3 games of 2 minutes. Round to Gemma. 3. Chess where the computer plays itself: Qwen 3:14, Gemma 1:21. Qwen's game: 482 moves in 5 minutes, all legal (checked with a chess library), every game a draw. Gemma's game freezes after 7 moves: two functions call each other forever. Round to Qwen. 4. Fix a cinema booking app with 7 hidden bugs and 12 tests: Qwen 0:34, Gemma 0:49. Both pass 12 of 12, tests untouched. A tie. Score: Qwen 2, Gemma 1, 1 tie. Time and the bill: across all 4 tests Gemma worked for under 4 minutes, Qwen for almost 10, and Qwen wrote about 3 times more tokens. Electricity at the US average price: Qwen 1.7 cents, Gemma half a cent. All 4 prompts are in the pinned comment. Chapters 0:00 The challenge 0:23 The setup 0:46 Test 1: 3D solar system 1:30 Test 2: Tetris 2:19 Test 3: Chess 3:09 Test 4: Fix the cinema app 3:35 Score, time and the bill 3:56 Verdict 4:08 Next episode More at kgptalkie.com