Tag
A technique called luce spark allows Qwen 35B-a3B MoE model to run on a 16GB GPU (like RTX 3090) by learning which experts are frequently used and streaming the rest from RAM, achieving ~100 tok/s without VRAM bottleneck.
A new open-source MoE model, gpt-oss-20b (21B total, 3.6B active), runs on only 1.8GB VRAM and achieves perfect scores on agentic coding tasks, outperforming other local models like Gemma and Qwen.