15-minute QLoRA finetune turns Qwen3.5-4B into a decision model playing Flappy Bird on 4GB VRAM

danielhanchen · x · 2026-10-13

A developer finetuned Qwen3.5-4B into a Jev-style decision model using UnslothAI, and the results are strikingly cheap and fast:

Trained on the LocalLLaMA/typed-decisions dataset with a decision head, using 4-bit QLoRA + bf16
Single A100, 15 minutes, 4 epochs; accuracy jumped from 30% to 80% (validation loss flattens after just 2 epochs)
Only 64.5M params (2.1% of the model) were trained; LoRA merged and quantized to Q4KM
Inference runs locally via llama.cpp on a single RTX 4060 playing Flappy Bird with 250ms latency in just 4GB VRAM — faster than most APIs tested

The recipe demonstrates that parameter-efficient finetuning of small models is a viable path for real-time on-device decision-making tasks.

More from Fun