Community Developer Shares Instruct Fine-Tuning Progress on Yandex AliceAI-80B-A3B Model

A community developer has completed initial post-training of an instruct version of Yandex's AliceAI-80B-A3B base model using a synthetically distilled dataset. The developer plans to release GGUF quants, public checkpoints, and a custom llama.cpp patch in the coming days. AliceAI-80B-A3B is an efficient open-weight Mixture-of-Experts (MoE) model featuring only 3 billion active parameters despite having 80 billion total parameters. Bringing GGUF quantization and llama.cpp support to this architecture enables open-source enthusiasts to run and evaluate a capable large-scale MoE model locally. The training ran over 48 hours using QLoRA across three 32GB V100 GPUs, powered by a custom training kernel and data distilled via the developer's SFTMill engine. Although loss curves fluctuated due to token count variations, average loss decreased steadily, with future plans including reinforcement learning and expanded fine-tuning.