Tag
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.