最上川 (@artorius)腾讯发布了首个 Diffusion 大语言模型 WeDLM-8B 中发帖

在数学推理任务中,相比经 vLLM 优化的 Qwen3-8B,速度提升 3–6 倍 
在大多数基准测试中,性能超越原始的 Qwen3-8B-Instruct
原生支持 KV Cache(兼容 FlashAttention、PagedAttention、CUDA Graphs)
[image]
[image]