AMD Open-Sources Instella-MoE-16B, a Fully Transparent Mixture-of-Experts LLM Trained on Its Own Instinct GPUs

AMD released the complete Instella-MoE-16B model under a research license, publishing weights from every training stage alongside data mixtures, training configs, and inference code. The model activates only 2.8 billion of its 16 billion parameters per token, offering researchers a rare fully documented MoE training run on non-Nvidia hardware.