AMD Releases Open Instella-MoE Model Trained on Instinct GPUs

AMD is expanding its generative AI strategy beyond Instinct accelerators by releasing models that researchers and companies can inspect, reproduce and adapt. Instella-MoE-16B-A3B is aimed primarily at academic and enterprise research teams, giving developers access to the components behind the system rather than only an inference endpoint or a limited set of model weights.

The chipmaker has released Instella-MoE-16B-A3B, a Mixture-of-Experts language model trained on AMD Instinct GPUs. The model contains 16 billion parameters in total but activates about 2.8 billion for each token, an approach designed to combine broader model capacity with lower computing demands. AMD also published the complete training data, configuration files and model weights, positioning the release as a fully open resource for AI research and development.