GLM 5.2

[reasoning]

Z.AI[text][open-weights]

Latest flagship with 1M context, stronger coding and long-horizon agentic reasoning

Interactive Architecture

Take GLM 5.2 apart, layer by layer. The 3D teardown below shows its 78 transformer layers (3 dense + 75 MoE), MLA attention, and expert routing, with values from the public model config.

78 layers (3 dense → 75 MoE)

MLA attention · 64Q heads

8+1 of 256 experts active

744B params · 40B active

Token embedding

6144-dim vectors over the same 154,880-token vocabulary as GLM-5. GLM-5.2 keeps the 744B backbone and stretches the context window to 1,048,576 tokens.

78 layers, IndexShare pattern

Still 3 dense-FFN layers then 75 MoE layers, plus one Multi-Token Prediction layer that Z.AI tuned for up to 20% longer speculative accept length. The new part is the indexer pattern: 21 layers run a full DSA indexer and the other 57 share one, in groups of four after the first three layers.

MLA + DSA with IndexShare

Multi-head Latent Attention with 64 heads and a 512-dim KV latent. A full layer runs the DSA indexer (32 heads, 128 dims) to pick the top 2048 cached tokens per query; the next three layers reuse that selection instead of scoring again. Z.AI reports 2.9x fewer per-token FLOPs at 1M context. RoPE theta also moves from 1e6 to 8e6 alongside the longer window.

256-expert MoE

Unchanged from GLM-5: a sigmoid router picks 8 of 256 routed experts per token, plus 1 shared expert (expert FFN 2048), so 40B of 744B parameters fire per token.

Output head

Final RMSNorm then an untied projection to 154,880 logits. With sparse attention and IndexShare, the 1M-token window is built for long-horizon agentic work.

Input

$1.40

per 1M tokens

Output

$4.40

per 1M tokens

Cached

$0.28

per 1M tokens

Context

1.0M tokens

Max Output

262K tokens

Released

2026-06-16

Features

[optional-reasoning][1m-tokens-context][engineering-grade-coding][long-horizon-planning][reasoning-with-think-tags][tool-calling]

Technical Specifications

LicenseMIT

Benchmarks

AIME 202699.2%

FrontierSWE74.4%

GPQA Diamond91.2%

Humanity's Last Exam40.5%

SWE-bench Pro62.1%

Terminal-Bench81.0%

The Lab

location: Beijing, Chinafounded: 2019backed by: Listed on HKEX (2513); early investors include Alibaba, Tencent, HongShan

Founded in 2019 as a Tsinghua University spinout, the company was known internationally as Zhipu AI until it rebranded as Z.ai in July 2025; it is headquartered in Beijing and has been listed on the Hong Kong Stock Exchange since January 2026. The lab develops the GLM (General Language Model) family, released under the MIT licence since GLM-4.5 in July 2025 and focused on coding and agentic tasks, alongside the GLM-V vision line and the AutoGLM phone agent.

more about Z.ai (Zhipu AI) →

How It Works

What vocabulary and embedding size does GLM 5.2 use?+

6144-dim vectors over the same 154,880-token vocabulary as GLM-5. GLM-5.2 keeps the 744B backbone and stretches the context window to 1,048,576 tokens.

How many transformer layers does GLM 5.2 have?+

What is Multi-head Latent Attention (MLA) in GLM 5.2?+

How does the Mixture of Experts in GLM 5.2 work?+

Unchanged from GLM-5: a sigmoid router picks 8 of 256 routed experts per token, plus 1 shared expert (expert FFN 2048), so 40B of 744B parameters fire per token.

How does GLM 5.2 produce output tokens?+

Final RMSNorm then an untied projection to 154,880 logits. With sparse attention and IndexShare, the 1M-token window is built for long-horizon agentic work.

scene values sourced from:HuggingFace config HuggingFace config https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm_moe_dsa/configuration_glm_moe_dsa.py paper https://sebastianraschka.com/llm-architecture-gallery/#card-glm-5-2

Model ID: glm-5.2

Interactive Architecture

Take GLM 5.2 apart, layer by layer. The 3D teardown below shows its 78 transformer layers (3 dense + 75 MoE), MLA attention, and expert routing, with values from the public model config.

78 layers (3 dense → 75 MoE)

MLA attention · 64Q heads

8+1 of 256 experts active

744B params · 40B active

Token embedding

6144-dim vectors over the same 154,880-token vocabulary as GLM-5. GLM-5.2 keeps the 744B backbone and stretches the context window to 1,048,576 tokens.

78 layers, IndexShare pattern

MLA + DSA with IndexShare

256-expert MoE

Unchanged from GLM-5: a sigmoid router picks 8 of 256 routed experts per token, plus 1 shared expert (expert FFN 2048), so 40B of 744B parameters fire per token.

Output head

Final RMSNorm then an untied projection to 154,880 logits. With sparse attention and IndexShare, the 1M-token window is built for long-horizon agentic work.

The Lab

location: Beijing, Chinafounded: 2019backed by: Listed on HKEX (2513); early investors include Alibaba, Tencent, HongShan

more about Z.ai (Zhipu AI) →

How It Works

What vocabulary and embedding size does GLM 5.2 use?+

6144-dim vectors over the same 154,880-token vocabulary as GLM-5. GLM-5.2 keeps the 744B backbone and stretches the context window to 1,048,576 tokens.

How many transformer layers does GLM 5.2 have?+

What is Multi-head Latent Attention (MLA) in GLM 5.2?+

How does the Mixture of Experts in GLM 5.2 work?+

Unchanged from GLM-5: a sigmoid router picks 8 of 256 routed experts per token, plus 1 shared expert (expert FFN 2048), so 40B of 744B parameters fire per token.

How does GLM 5.2 produce output tokens?+

Final RMSNorm then an untied projection to 154,880 logits. With sparse attention and IndexShare, the 1M-token window is built for long-horizon agentic work.