Qwen3.8-2.4T Is Here: Alibaba 2.4 Trillion Parameter MoE Model Just Changed the Open Source AI Landscape

Alibaba's Qwen3.8 model goes live on Nvidia's GB300, hits 4,000 tokens per second

Alibaba released Qwen3.8-Max (2.4T sparse-MoE), hitting 4,000 tokens/sec on Nvidia GB300 and scoring 86.1 OSWorld-Verified, outperforming Claude/GPT variants. At $2 per 1M input tokens with open weights, Qwen3.8 pressures OpenAI/Claude pricing while fragmenting closed-model dominance for enterprises.