Kisoku 1.6B, trained from scratch by one person, runs on our CPU-only box from a 1.7 GB download
Kisoku 1.6B from 0ARCH was trained from scratch by one person, Joseph Rodriguez, and is free under Apache 2.0; the recommended chat download is 1.7 GB and runs in Ollama. The chat model is a preview and, per its maker, can state wrong answers confidently. Our test on a CPU-only box with 8 cores and no GPU: a correct Python string-reverse function and "Canberra" for the capital of Australia, at about 40 tokens/s using about 2.5 GB of RAM (numbers measured on our test box, not typical performance).
Try it
ollama create kisoku -f Modelfile && ollama run kisokuInstall Ollama from ollama.com first. In a new folder, download the 1.7 GB chat model and its Modelfile: curl -LO https://huggingface.co/0arch-io/kisoku-1.6b-gguf/resolve/main/kisoku-1.6b-chat-sft009-Q8_0.gguf and curl -LO https://huggingface.co/0arch-io/kisoku-1.6b-gguf/resolve/main/Modelfile. We ran it with --verbose added on our CPU-only test box (8 cores, no GPU); speed and memory there are not typical performance.