Unsloth is the reason consumer GPUs can train real models — not toy quantized models, actual 8B-70B instruction tuning.
The trick: custom Triton kernels for the attention and MLP layers, plus memory-optimized optimizer states. You get 2x speed and 70% less VRAM than a naive Hugging Face Trainer setup.
Practical result: fine-tune Llama 3.1 8B on a single 24GB RTX 4090. Fine-tune 70B with 4-bit quantization on a 48GB A6000 or dual RTX 4090s.
The API: beautifully simple. from unsloth import FastLanguageModel, load in 4-bit, apply LoRA, train. Fifteen lines of Python and you're fine-tuning.
The ecosystem: Unsloth Studio — a local UI that wraps everything. Upload dataset, pick model, click Train, download GGUF.
What it doesn't do: full pretraining from scratch (that needs a datacenter) and extremely exotic architectures.
Who this is for: anyone who wants a custom model without renting a cluster or a compliance headache. The fully local path finally works.
Our results: we tuned Qwen2.5-Coder-7B on our internal code review corpus. It now catches the naming inconsistencies our linter misses — and it runs on one laptop.