Something strange happened in open-weight AI this week. Z.ai and Alibaba’s Qwen team both shipped frontier-scale models within 24 hours of each other, and the architectures are so similar they could have come from the same internal memo. GLM-5.3-Flash runs 320B parameters with 18B active, while Qwen3.8-Flash-Next runs 125B with 6B active. The configs match on nearly every consequential design decision.
Both models use a 3:1 hybrid of linear and full attention layers. GLM stacks 34 linear-attention layers against 11 full-attention layers; Qwen runs a repeating block of 3 Gated DeltaNet layers plus 1 Qwen Sparse Attention layer. The linear layers compress history into a fixed-size recurrent state, eliminating the growing KV cache problem that makes long contexts expensive. The full-attention layers handle precise long-range retrieval. Different gating granularity — GLM uses Kimi Delta Attention with per-channel decay, Qwen uses per-head gating — but the delta-rule family is identical.
The shared recipe doesn’t stop there. Both models attach a learned indexer that compresses context 4x before scoring, then keeps exactly 2048 tokens for full attention. GLM’s lightning indexer selects top-2048 directly; Qwen scores 4-token blocks and keeps the top 512 blocks, which lands on the same number. Qwen credits this approach with up to 7.6x prefill speedups at 1M tokens. Both models also abandon the single residual stream, widening it into 4 gated branches. GLM adopted DeepSeek’s Manifold-Constrained Hyper-Connections; Qwen wrote its own Gated Residual variant that drops the branch-mixing step and enables FP8 residual storage.
What makes this convergence notable is that neither team copied the other — they designed independently. That means the field has landed on something close to a consensus architecture for efficient frontier models. The one lab dissenting from this recipe matters more than the agreements. When independent teams reach identical conclusions about linear attention ratios, indexer budgets, and residual stream design, it stops being speculation and starts being the new default. The open-weight ecosystem just standardized itself in real time.
