FrontierAugust 1, 2026via MarkTechPost

AMD Releases Instella-MoE-16B-A3B: A Fully Open Mixture-of-Experts LLM With 2.8B Active Parameters Trained On Instinct GPUs

Why it matters

AMD is using open-weight model releases to establish Instinct GPUs as a credible training platform and compete with Nvidia-centric model ecosystems. Full transparency on training (weights, data, configs) is rare and shifts how practitioners evaluate hardware-model fit.

Key signals

  • Model: Instella-MoE-16B-A3B (16B total parameters, 2.8B active per token)
  • Architecture: Mixture-of-Experts with Gated MLA and FarSkip-Collective
  • Hardware: Trained on AMD Instinct MI300X and MI325X GPUs
  • Open release: weights from every training stage, data mixtures, configs, and inference code published
  • Significance: demonstrates AMD GPU capability for frontier model training; full transparency rare in open-weight releases

The hook

AMD drops a fully open MoE model trained on its own silicon—16B params, 2.8B active, weights-to-inference published.

AMD released Instella-MoE-16B-A3B, a fully open Mixture-of-Experts language model trained from scratch on Instinct MI300X and MI325X GPUs. It holds 16B total parameters but activates only 2.8B per token, using Gated MLA and FarSkip-Collective. AMD published weights from every training stage, plus da

The week's key stories, every Friday.

For practitioners and enthusiasts — free, in your inbox.

Free forever. No spam.