HobbyLM-1B Instruct

A small sparse mixture-of-experts model (1.04B total, ~305M active per token) trained from scratch on 100B tokens and instruction-tuned by Fuel Labs.

Float32 weights, greedy decoding, 256-token answers, 3840-token prompt budget. Answers are often wrong or repetitive and the model has no safety tuning — it is a research demo.

Model card · Base model · GGUF