AMD launches Instella-MoE-16B-A3B, trained entirely on its own GPUs

AMD has launched its Instella-MoE-16B-A3B AI model, a significant development as it was trained entirely on AMD's own GPUs, specifically the Instinct MI300X and MI325X, without relying on Nvidia hardware or software like CUDA. This 16-billion parameter model utilizes a Mixture-of-Experts (MoE) architecture, with only 2.8 billion parameters actively used per inference, enabling faster processing. AMD has also made the training weights, code, and data recipes fully open-source, allowing for community inspection and further development, though the model weights are restricted to research purposes under the ResearchRAIL License. AI

IMPACT This release challenges Nvidia's dominance in AI training by demonstrating a viable alternative ecosystem, potentially lowering costs and increasing accessibility for researchers and developers.

RANK_REASON The cluster describes a new AI model release from a major hardware vendor (AMD) that showcases their own training infrastructure, positioning it as a viable alternative to the dominant player (Nvidia). [lever_c_demoted from frontier_release: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• AMD
• CUDA
• Instella-MoE-16B-A3B
• Instinct MI300X
• Instinct MI325X
• MIT License
• Mixture-of-Experts (MoE)
• Nvidia
• ResearchRAIL License
• ROCm

• model release
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →