GLM-5.3-Flash and Qwen3.8-Flash-Next converge on hybrid linear attention, sparse 2048-token indexers, and four residual streams.