@build.with.alan: Comment SOUP and I’ll send you the repo. 👇 An 8B model. A 4GB laptop GPU. 3.3 gigs of VRAM. 🍜 Alpamys Makazhan is a solo founder from Kazakhstan. He wanted to fine-tune models that didn’t fit on his laptop — so he built a tool that made them fit. It’s called Soup. And it kills the exact reason most people never fine-tune anything: renting cloud GPUs, writing training scripts, debugging for days, then giving up and using someone else’s model. Soup collapses all of it into one YAML file and one command. Describe what you want the model to do, point it at your data, run soup train. Batch size, quantization, LoRA settings — handled. The real trick is layer streaming. A full 8B model like Llama normally needs 16GB of VRAM. Soup quantizes the frozen base to 4-bit, keeps it in system RAM, and feeds the GPU one decoder layer at a time. Train it, swap it, next. Peak VRAM — 3.3GB. 119 tokens a second. On a 4GB RTX 3050. Then someone lent him 8 H100s… and the laptop numbers held up. Front page of Hacker News. Apache 2.0. Completely free. 🔓 Constraints don’t kill good engineering. They design it.

Alan Brown l AI & Robotics

Open In TikTok:

Region: SG

Sunday 16 August 2026 10:15:15 GMT