AMD releases Instella-MoE-16B-A3B, a fully open MoE LLM trained on Instinct GPUs

AMD has released Instella-MoE-16B-A3B, a fully open Mixture-of-Experts (MoE) large language model trained on its own Instinct GPUs. The model, reported by Marktechpost, features 2.8 billion active parameters and is the latest addition to AMD's Instella family of open models.

The naming gives away the architecture: the 16B figure points to roughly 16 billion total parameters, while the A3B designation signals the number of parameters activated during each inference step, put at 2.8 billion in the report. MoE designs work by routing each token through only a fraction of the expert sub-networks, so the model keeps a large total capacity while cutting per-query compute costs.

The fully open nature of the release is its most significant aspect. Developers can download the weights, fine-tune the model, and self-host it themselves, without sending data through a vendor's API — something closed models do not allow. The move continues the open-source approach AMD has taken with the Instella series.

There is a clear hardware signal here as well. Training the model on AMD Instinct GPUs is a statement that AMD's accelerators can handle not just inference but also the training workloads behind frontier open models, strengthening the case for an end-to-end AMD ecosystem from training to deployment.

Low-active-parameter MoE models have become a favored direction in the open-source community for efficient inference. With 16B total parameters and 2.8B active, Instella-MoE-16B-A3B enters that race as a new self-hosted option for developers and organizations with limited compute budgets.

What to watch next: the actual availability of the weights, the model's results on mainstream benchmarks, and how it compares with other open MoE models of a similar class. The maturity of AMD's ROCm software stack for this model will also play a large role in how widely it gets adopted in production.