GLM-5.3-Flash vs Qwen3.8-Flash-Next: Two Chinese AI Labs Independently Converge on the Same Model Architecture

Z.ai and Alibaba’s Qwen team released frontier open-weight models within a day of each other, and the surprise is not just that they both went big. It’s that their blueprints line up so closely. GLM-5.3-Flash is a 320B-parameter multimodal MoE model with 18B active parameters. Qwen3.8-Flash-Next is a 125B model with 6B active parameters, and it serves as a preview of Qwen4’s architecture.

The cleanest overlap is the attention stack. GLM-5.3-Flash uses 34 linear-attention layers and 11 full-attention layers. Qwen3.8-Flash-Next uses repeating blocks of 3 Gated DeltaNet layers and 1 Qwen Sparse Attention layer. Different names, same 3:1 split. The cheap layers keep a fixed recurrent state instead of a growing KV cache, so token cost stays flat as context grows. The precise layers still handle long-range retrieval.

Both teams also landed on the same trick for sparse attention: compress the history, score the compressed chunks, and keep only 2048 tokens. GLM does it with a 32-head lightning indexer and an IndexPool compression step that folds four key vectors into one before scoring. Qwen scores 4-token blocks and keeps the top 512 blocks, which also comes out to 2048 tokens. The result is the same basic bargain: a much smaller attention budget at long context lengths.

They made the same architectural move in the residual path too. Both models widen the single transformer residual stream into four gated branches. Z.ai uses Manifold-Constrained Hyper-Connections. Qwen uses Gated Residual, and the team says its own ablation found it roughly equal in quality to mHC. Both models are also trained with Muon, and both split fused projection matrices before orthogonalization.

The one sharp disagreement is positional encoding. GLM-5.3-Flash drops RoPE in its sparse layers entirely. Qwen tested NoPE as well, saw no real pre-training difference, and then ran into a nastier problem after post-training: the model often failed to stop generating. That is the kind of detail model cards rarely shout about, but it matters more than another glossy benchmark slide. MiniMax is the bigger dissenting voice, though, saying linear attention hurt multi-hop reasoning in its own scaling tests and shipping sparse softmax instead.

The broad pattern is hard to miss. DeepSeek, Moonshot, Z.ai, and Qwen are all converging on a shared Chinese open-model recipe: sparse indexing, 2048-token attention budgets, four-branch residuals, Muon. The argument now is not whether the parts work in isolation. It’s whether this is the new default architecture, or just a very well-funded coincidence.