you can build something like Bonsai 2 for ~$43k. you too could be like @PrismML, you just need one on-demand B300 node for ~4 weeks. i wrote AutoAWQ (the quantization library behind 10k models) from scratch. here's how i'd reproduce a great ternary model: 1. start with a hadamard fold into every projection (you can use block-1024 like prism) 2. apply CAT-Q quantization (TernaryQuench reproduces it). scale the dataset so it takes ~3 days with 10-20M tokens 3. first round of distillation: KL…