GLM 5.1

[reasoning]

Z.AI[text][open-weights]

Flagship coding model with extended autonomous capability for long-horizon, engineering-grade tasks

Interactive Architecture

Take GLM 5.1 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-5.1 reuses the GLM-5 backbone unchanged.

78 layers (3 dense prefix)

The GLM-5 stack as-is: 3 dense-FFN layers followed by 75 MoE layers, with one Multi-Token Prediction layer on top for speculative decoding. Every architecture field in config.json matches GLM-5.

MLA + DeepSeek Sparse Attention

Multi-head Latent Attention with 64 heads, a 2048-dim query LoRA and a 512-dim KV latent. A DSA indexer (32 heads, 128 dims) keeps the top 2048 cached tokens for each query, so long agentic sessions stay affordable.

256-expert MoE

A sigmoid router picks 8 of 256 routed experts per token, plus 1 shared expert (expert FFN 2048). 40B of 744B parameters fire per token.

Output head

Final RMSNorm then an untied projection to 154,880 logits. What changed is post-training: GLM-5.1 is tuned to stay effective over hundreds of rounds and thousands of tool calls in agentic coding.

Input

$0.95

per 1M tokens

Output

$3.15

per 1M tokens

Cached

$0.47

per 1M tokens

Context

203K tokens

Max Output

66K tokens

Released

2026-04-07

Features

[optional-reasoning][autonomous-8hr+-coding-sessions][engineering-grade-output][long-horizon-planning][reasoning-with-think-tags][tool-calling]

Technical Specifications

LicenseMIT

Architecture

ArchitectureMixture of Experts (MoE)

Parameters744B total / 44B active (estimated)

Resources

Demo

Benchmarks

AIME 202695.3%

BrowseComp68.0%

CyberGym68.7%

GPQA Diamond86.2%

Humanity's Last Exam31.0%

MCP-Atlas71.8%

SWE-bench Pro58.4%

Terminal-Bench63.5%

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 5.1 use?+

6144-dim vectors over the same 154,880-token vocabulary as GLM-5. GLM-5.1 reuses the GLM-5 backbone unchanged.

How many transformer layers does GLM 5.1 have?+

The GLM-5 stack as-is: 3 dense-FFN layers followed by 75 MoE layers, with one Multi-Token Prediction layer on top for speculative decoding. Every architecture field in config.json matches GLM-5.

What is Multi-head Latent Attention (MLA) in GLM 5.1?+

How does the Mixture of Experts in GLM 5.1 work?+

A sigmoid router picks 8 of 256 routed experts per token, plus 1 shared expert (expert FFN 2048). 40B of 744B parameters fire per token.

How does GLM 5.1 produce output tokens?+

Final RMSNorm then an untied projection to 154,880 logits. What changed is post-training: GLM-5.1 is tuned to stay effective over hundreds of rounds and thousands of tool calls in agentic coding.

scene values sourced from:HuggingFace config HuggingFace config paper https://sebastianraschka.com/llm-architecture-gallery/#card-glm-5-1

Model ID: glm-5.1

Interactive Architecture

Take GLM 5.1 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-5.1 reuses the GLM-5 backbone unchanged.

78 layers (3 dense prefix)

The GLM-5 stack as-is: 3 dense-FFN layers followed by 75 MoE layers, with one Multi-Token Prediction layer on top for speculative decoding. Every architecture field in config.json matches GLM-5.

MLA + DeepSeek Sparse Attention

256-expert MoE

A sigmoid router picks 8 of 256 routed experts per token, plus 1 shared expert (expert FFN 2048). 40B of 744B parameters fire per token.

Output head

Final RMSNorm then an untied projection to 154,880 logits. What changed is post-training: GLM-5.1 is tuned to stay effective over hundreds of rounds and thousands of tool calls in agentic coding.

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 5.1 use?+

6144-dim vectors over the same 154,880-token vocabulary as GLM-5. GLM-5.1 reuses the GLM-5 backbone unchanged.

How many transformer layers does GLM 5.1 have?+

The GLM-5 stack as-is: 3 dense-FFN layers followed by 75 MoE layers, with one Multi-Token Prediction layer on top for speculative decoding. Every architecture field in config.json matches GLM-5.

What is Multi-head Latent Attention (MLA) in GLM 5.1?+

How does the Mixture of Experts in GLM 5.1 work?+

A sigmoid router picks 8 of 256 routed experts per token, plus 1 shared expert (expert FFN 2048). 40B of 744B parameters fire per token.

How does GLM 5.1 produce output tokens?+

Final RMSNorm then an untied projection to 154,880 logits. What changed is post-training: GLM-5.1 is tuned to stay effective over hundreds of rounds and thousands of tool calls in agentic coding.