Unsloth trains local decision models on 3GB VRAM, lifting accuracy from 30% to 78%
evilsocket · x · 2026-10-08
Unsloth now lets you turn open LLMs like Qwen, Gemma and Llama into local decision models with just 3-4GB of VRAM.
- Fine-tuning with a Clef head raised Qwen3.5 0.8B's aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks
- One epoch of Unsloth + LoRA (r=64) fine-tuning boosted downstream accuracy from 30-37% to 78%
- Open-source repo, guides and notebooks available, plus Unsloth Desktop
Related event: Unsloth Enables Local Training of Decision Models with Under 4GB VRAM(2 posts)→
More from coding & agent
- Dev dumps 6x-Sol over quality, burns $200 Astra sub in half a day, moves to Claude Code — Late_Change5029 · 2026-10-08
- This Solo Dev Runs His Entire Business From One Obsidian Vault With AI — dSebastien · 2026-10-08
- Cognizant to Hire 1,500 US Graduates; Devin Cut Freight Firm's Rebuild Costs 37% — shashib · 2026-10-08
- Cisco Brings Claude Managed Agents to Webex With Its Own Governance Layer — shashib · 2026-10-08
- Dev finds 6.1 Sol surprisingly good at designing native iOS apps — Dimillian · 2026-10-08
- Anthropic test of 52 devs: AI-assisted group scored 50% vs 67% without, error-finding worst — AlexTensor · 2026-10-08