Unsloth Enables Local Decision Model Training on 3GB VRAM

Unsloth has released tools enabling local fine-tuning of models such as Llama, Gemma, and Qwen into decision-making agents on consumer hardware with as little as 3GB VRAM. The system uses a Clef head and LoRA for one epoch of training via the Unsloth Desktop application. In internal tests, the Qwen3.5 0.8B model's aggregate accuracy increased from 20.7% to 74.3% across three decision benchmarks.