Actually put Gemma 4 26B to work on something real: had it extract trading signals from 2,400 earnings calls. One worked. One almost fooled me.

Everyone posts benchmarks and arena scores. I wanted to see if a local model could do something that makes actual money. So I took my Gemma 4 26B (IQ4_XS quant, running on a single 4090) and gave it a job: read 2,400 earnings call transcripts from the last 3 years and find language patterns that predict how the stock moves in the 5 days after. Fine-tuned on about 800 labeled transcripts. The labels were simple: did the stock beat or miss its sector over the next week. Model's job wasn't price prediction.