GLM-5.2
A later GLM line member that extends sparse-attention mixture-of-experts architecture with a solid 1M-token context and long-horizon coding emphasis.
GLM-5.2 matters because it keeps the GLM sparse-attention, mixture-of-experts stack but pushes the line toward long-horizon coding work. The official release blog emphasizes a stable 1M-token context, IndexShare efficiency for sparse attention, and training aimed at sustained agentic engineering rather than one-shot answers. The Z.ai GitHub download table lists the checkpoint at 744 billion total parameters with about 40 billion active per token.
At a glance
- Family
- glm
- Source type
- Open weights
- Released
- June 2026
- Authors
- Z.ai
- Modalities
- Text
- Parameter count
- 744 billion total parameters
- Context length
- 1,048,576 tokens
- Precision
- bf16
What It Is
GLM-5.2 is an open-weights text model from Z.ai, released on 2026-06-17 in the GLM-5.2 blog post. It is the later GLM line member that builds on the same sparse mixture-of-experts transformer family as GLM-5 while adding a 1M-token context window under an MIT open-source license. The official GitHub model download table lists it at 744 billion total parameters and about 40 billion active parameters.Inputs And Outputs
GLM-5.2 is a text-in, text-out model. The public release emphasizes long coding-agent trajectories, tool use, and multi-step engineering work where the model must keep quality across very long prompts instead of degrading once context grows.Architecture
GLM-5.2 keeps the decoder-only sparse MoE skeleton from the GLM line. The release blog highlights IndexShare, which reuses a lightweight indexer across every four sparse-attention layers to cut indexer cost at 1M context, plus improved multi-token prediction for speculative decoding. The named building blocks still include DSA-style sparse attention, mixture-of-experts routing, multi-head latent attention, RMSNorm, and RoPE.Output
ProbabilitiesSoftmaxOutput
LinearTransformer architectureAttention sublayer containerSparse expert sublayer containerN×RMSNormSparse
Attention
(DSA)RMSNormMixture
Of ExpertsEmbeddingRoPEInput
EmbeddingInput
TokensSoftmax to Output
ProbabilitiesOutput
Linear to SoftmaxRMSNorm to Output
LinearMixture
Of Experts to RMSNormRMSNorm to Mixture
Of ExpertsRMSNorm to RMSNormSparse
Attention
(DSA) to RMSNormEmbedding to Sparse
Attention
(DSA)Embedding to RMSNormInput
Embedding to EmbeddingRoPE to EmbeddingInput
Tokens to Input
Embedding
Important Modules
These modules explain why GLM-5.2 stays in the same modern long-context MoE lane as GLM-5 rather than becoming a separate architecture family. Sparse attention and mixture-of-experts carry the scaling story, while latent attention, RMSNorm, and RoPE support stable training and very long prompts.Training
The GLM-5.2 blog frames post-training around long-horizon agentic reinforcement learning, on-policy distillation from multiple expert models, and anti-hack safeguards for coding-agent rollouts. The public materials describe expanded 1M-context training for coding-agent scenarios and open-weight release in bf16.Training regimes
Linked papers
- GLM-5