AMD RELEASES INSTELLA-MOE-16B-A3B: A FULLY OPEN MIXTURE-OF-EXPERTS LLM WITH 2.8B ACTIVE PARAMETERS TRAINED ON INSTINCT GPUS

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.

MarkTechPost reports that aMD released Instella-MoE-16B-A3B, a fully open Mixture-of-Experts language model trained from scratch on Instinct MI300X and MI325X GPUs.

The report adds: 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 data mixtures, configs, and inference code.

This report is covered across the current DAMMNEWS feeds. Technology stories often mix a product announcement with claims about availability, performance or impact. Those parts should be checked separately.

This is a developing report. DAMMNEWS will surface related coverage as closely matched reporting enters the live cache.

DAMMNEWS RSS BRIEFING — generated locally from the available MarkTechPost headline and RSS excerpt. It is not a summary of the full source article.