AMD unveils its in-house AI model “Instella-MoE,” trained on its own GPUs, a small model that outperforms Gemma-4-E4B

AMD has released Instella-MoE, an MoE language model trained on its Instinct MI300X and MI325X GPUs.

The company is offering multiple versions, from pretraining to reinforcement-learning models, along with training code for free on Hugging Face and GitHub.

Despite using fewer active parameters, it delivered results that beat Gemma-4-E4B-it, highlighting its strength among small open models.

As a fully open model trained on AMD hardware and software such as ROCm, it strengthens AMD’s position as an open AI development platform.