Unsloth fine-tunes Qwen3.5 0.8B on 4GB VRAM, lifting decision accuracy from 20.7% to 74.3%
kalyan_kpl · x · 2026-10-08
Unsloth AI announced you can now train your own local decision model: fine-tuning Qwen3.5 0.8B with a Clef head and LoRA (r=64) for one epoch raised aggregate accuracy across 3 decision benchmarks from 20.7% to 74.3%, with downstream accuracy up from 30–37% to 78% — all on just 4GB of VRAM.
- Method: Unsloth + LoRA (r=64), single epoch
- Open-source repo (Apache-2.0/AGPL-3.0, 77.3k GitHub stars) supports GGUF, MLX, Qwen, DeepSeek, Gemma, FLUX and more
- Official guide and notebooks included for turning any small LLM into a decision model
More from Models
- ThePrimeagen benchmarks OpenAI's new decision model: faster and more accurate on GUI tasks — prd_008 · 2026-10-08
- AI2's Bolmo tackles the 'token tax' hitting Global South scripts — Kyle_L_Wiggers · 2026-10-08
- Grok bot goes proactive on X, now suggests what you should do next — Daniel_Farinax · 2026-10-08
- OpenAI and Cloudflare launch decision APIs — tested at 355 decisions, TypeSafe's Jev still decides more — PawelHuryn · 2026-10-08
- Gary Marcus: OpenAI's vague math report 'would never pass peer review' — Tao responds too — Gary Marcus · 2026-10-08
- d1-3B runs on a MacBook: hands-on demos show Liquid AI's open model in action — JosephJacks_ · 2026-10-08