GLM-5 moved Z.ai's flagship from GLM-4.5's grouped-query attention to multi-head latent attention with DeepSeek Sparse Attention (DSA) in every layer. Each of the 78 blocks caches a 512-dim KV latent plus a 64-dim rotary key, and a lightning indexer (32 heads × 128 dims) picks the 2,048 past tokens each query attends to. DSA was added by continued training after mid-training (a 1,000-step indexer warm-up, then 20B tokens of sparse adaptation). In Z.ai's own ablation on a 9B model, sliding-window and Gated DeltaNet hybrids that kept half their layers as full attention still lost 5.7–30.4 points on RULER at 128K, which is the case it makes for token-level sparsity over layer-level hybrids.
Two MLA changes are specific to GLM-5. Under Muon, the 576-dim latent at first trailed GQA-8; orthogonalizing each head's slice of the up-projections separately (Muon Split) closed the gap and kept attention logits stable without clipping. And because DeepSeek-V3's head count was tuned to the H800 roofline, GLM-5 uses fewer, wider heads (MLA-256: 64 heads, q/k head dim 256 = 192 + 64 rotary, v head dim 256) to cut decode compute. The FFN is DeepSeek-style: 3 dense layers, then 75 layers of 256 routed experts (top-8, sigmoid scores, bias-corrected selection) plus one shared expert. A single MTP layer is reused for 3 prediction steps.
GLM-5.2 (June 2026) keeps the 744B-A40B backbone and changes three things. IndexShare runs the indexer in only 21 of 78 layers (the first three, then every fourth), and the other 57 reuse the top-k of the preceding indexer layer (2.9× fewer per-token FLOPs at 1M tokens). The RoPE base rises from 1M to 8M and the window from 202,752 to 1,048,576 tokens. The MTP layer reuses the first draft step's indices and KV in later steps and is trained with rejection sampling and a total-variation loss. GLM-5.1 shipped the GLM-5 config unchanged, and GLM-5.3 reuses the GLM-5.2 base model.