AMD releases Instella-MoE-16B-A3B, an open-source MoE model
16B total parameters, only 3B active per token—AMD enters the open-source AI race
Deep Dive
AMD's Instella-MoE-16B-A3B model has been uploaded to HuggingFace, spotted by a Reddit user who says it's good to see AMD joining the open source AI game.
Key Points
- 16B total parameters with MoE sparse activation of only 3B per token for efficient inference.
- First open-source LLM from AMD, hosted on HuggingFace with permissive licensing.
- Optimized for ROCm stack, aiming to reduce dependency on NVIDIA CUDA infrastructure.
Why It Matters
AMD democratizes MoE AI, giving devs a hardware-agnostic, efficient alternative to dense models.