AMD Launches Instella-MoE: Its First Fully Open-Source MoE Model

BanghuaZ · x · 2026-07-30

AMD has announced Instella-MoE, its first fully open-source Mixture-of-Experts (MoE) language model. The model features 16B total parameters with only 2.8B active parameters per token.

Trained from scratch on AMD Instinct MI300X and MI325X GPUs, Instella-MoE combines a sparsely activated MoE design with architectural innovations such as Gated Multi-head Latent Attention (Gated MLA) and FarSkip-Collective. It achieves state-of-the-art performance among fully open models at its scale.

The release provides the complete end-to-end training details. Post-training RL via the Miles framework boosted IFEval scores from 77.1 to 83.7, while SGLang serving delivers up to 39.2% lower Time-To-First-Token (TTFT) under expert parallelism.

Related event: AMD Releases Fully Open-Source MoE Model Instella(2 posts)→

Original post →

More from Models

Models channel →