Can a 12B Model Code on a Laptop? I Ran 10 Brutal Tests…

I ran Google's Gemma 4 12B — fully local on an RTX 4060 laptop (8GB) — through 10 hard, one-shot coding challenges. One prompt each, no retries, no fixing its mistakes. Whatever it wrote, I ran immediately.

Result: 3 passed, 7 failed

Passes: Conway's Game of Life, a Web Audio drum machine, an SVG fox from text.

Fails — and they're wild: a model that clearly understands what to build but fumbles the typing — declaring a function twice (Mandelbrot), typing THeRE instead of THREE (Three.js city), dropping the dot in getElementById (space shooter), Tetris where no piece spawns, a piano that forgot its keys, a WebGL raymarcher that won't compile, and a bug-fix task where it fixed two and added a third.

The honest takeaway

Almost every failure was a one-character typo a coding agent would fix on the first retry. Raw one-shot? 3/10. Inside an agent loop? Much higher.

Everything captured live — real OBS recordings, real browser, real errors. Model: Gemma 4 12B IT Q4_K_M (Unsloth GGUF) · llama-server · RTX 4060 8GB.

All 10 tests in the video above. Subscribe for more real local-LLM coding tests.