Open-Weight Model Race Heats Up

Alibaba's Qwen3.8-Max (2.4T) claims open-source lead with open weights next week. DeepSeek V4 Flash (13B active MoE) offers low-cost alternative. Meta released Muse Glimmer, a 30B dense open-weight edge model for local agentic AI, optimized for long-running tasks on NVIDIA hardware (RTX 5090, DGX Spark, Jetson). Dense architecture avoids MoE overhead. Sakana Namazu, LLaDA MoE v2, Shieldstral, Muse Spark 1.2 crack top 5 on Vals Index.