AMD is betting that openness will make its Instinct accelerators harder to ignore in the AI training race. The company released Instella-MoE-16B-A3B this week, an open mixture-of-experts model with 16B total parameters and 2.8B active per token, trained from scratch on MI300X and MI325X GPUs.
What sets the release apart is the completeness of the disclosure. AMD published checkpoints from every training stage, plus the data mixtures, training configs and serving code, so outside researchers can reproduce the full recipe rather than just fine-tune a final model.
Two design choices drive the efficiency story. Gated Multi-head Latent Attention adds a learned gate to the attention output, and FarSkip-Collective overlaps communication with computation in expert-parallel training, yielding a 12.7% pre-training speedup and up to 39.2% faster time to first token in serving.
AMD’s base checkpoint posts a 76.7 average on standard tests, which it calls the best result among fully open models, just ahead of Moonlight-16B-A3B’s 76.2 and behind Qwen3.5-4B-Base’s 79.5. Post-training then lifts instruction-following on IFEval from 77.08 to 83.70.
The weights carry a research-only ResearchRAIL license, so enterprises eyeing production use will likely start with the MIT-licensed training code on GitHub. Memory needs are modest, around 32GB in BF16, keeping the model within range of a single high-memory accelerator.
