GLM-5.2

About

GLM-5.2 (zai-org/GLM-5.2) is Z.ai's flagship model for long-horizon tasks - a large sparse Mixture-of-Experts LLM (~744-753B total parameters, ~40B active per token) that for the first time delivers long-horizon capability on a solid 1M-token context, released June 16, 2026 under the MIT License (no regional limits).

Its headline improvements over GLM-5.1: a stable 1M context that sustains long-horizon work; advanced coding with flexible thinking effort levels to balance performance and latency; and an improved architecture built around IndexShare - reusing the same sparse-attention indexer across every four attention layers (21 full + 57 shared indexers in the 78-layer stack, index top-k 2048 with 32 index heads of dim 128), which cuts per-token FLOPs by 2.9x at 1M context. The backbone (GlmMoeDsa) pairs DeepSeek-Sparse-Attention-style indexing with MLA-style compressed KV (kv-LoRA rank 512, 64 heads of dim 192), 3 dense + 75 sparse MoE layers, 256 routed experts + 1 shared with 8 routed per token (expert FFN 2048, sigmoid scoring, noaux-tc), hidden size 6144, vocab 155K, and an improved Multi-Token Prediction (MTP) layer for speculative decoding with up to 20% longer acceptance length.

GLM-5.2 leads open-weights models on the Artificial Analysis Intelligence Index (v4.1, 11 points above GLM-5.1), with strong agentic/coding results (SWE-Bench Pro, NL2Repo, DeepSWE, HLE-with-tools at up to 400K context).

Training Data Trained as part of GLM-5 series with long-horizon task focus