Model releases & benchmarks
Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency
Qwen3.8-Flash-Next pairs compressed sequence memory with selective retrieval, and adds gated residual pathways and N-gram lookup capacity to target long-context efficiency.
Why it matters: For long-context and tool-driven workloads, the design aims to reduce attention and memory costs while retaining retrieval and model capacity. The post also gives a concrete distinction between kernel speedups and serving throughput, which is reported under a 90% Prefix Cache hit rate.