GLM 4.7 Flash

[reasoning]

Z.AI[text][open-weights]

Efficient 30B model optimised for agentic coding and long-horizon planning

Interactive Architecture

Take GLM 4.7 Flash apart, layer by layer. The 3D teardown below shows its 47 transformer layers (1 dense + 46 MoE), MLA attention, and expert routing, with values from the public model config.

47 layers (1 dense → 46 MoE)

MLA attention · 20Q heads

4+1 of 64 experts active

30B params · 3B active

Token embedding

2048-dim vectors over a 154,880-token vocabulary. The Flash tier shrinks the hidden dim to 2048 for fast, cheap inference.

47 layers (1 dense prefix)

Pre-norm blocks with the first layer dense and the rest MoE, plus a 1-layer MTP head. Roughly half the depth of GLM-4.7.

Multi-head Latent Attention

20 heads over a 512-dim KV latent (q_lora_rank 768). Unlike the GQA flagship, the Flash tier uses MLA-style compression to shrink the KV cache further.

64-expert MoE

A router scores 64 experts and activates 4 plus 1 shared, spending just 3B of 30B per token. The Lite routing keeps latency low.

Output head

Final RMSNorm then 154,880 logits. Flash is the cost-optimised GLM-4 tier for high-volume serving.

Input

$0.07

per 1M tokens

Output

$0.40

per 1M tokens

Context

200K tokens

Max Output

131K tokens

Released

2026-01-19

Features

[optional-reasoning][agentic-coding][long-horizon-planning][tool-collaboration][cost-effective]

Technical Specifications

LicenseMIT

Architecture

ArchitectureMixture of Experts (MoE)

Parameters30B total / ~3B active

Resources

HuggingFace Demo

Benchmarks

AIME 202591.6%

BrowseComp42.8%

GPQA Diamond75.2%

Humanity's Last Exam14.4%

LiveCodeBench64.0%

SWE-bench Verified59.2%

Tau2-bench79.5%

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 4.7 Flash use?+

2048-dim vectors over a 154,880-token vocabulary. The Flash tier shrinks the hidden dim to 2048 for fast, cheap inference.

How many transformer layers does GLM 4.7 Flash have?+

Pre-norm blocks with the first layer dense and the rest MoE, plus a 1-layer MTP head. Roughly half the depth of GLM-4.7.

What is Multi-head Latent Attention (MLA) in GLM 4.7 Flash?+

20 heads over a 512-dim KV latent (q_lora_rank 768). Unlike the GQA flagship, the Flash tier uses MLA-style compression to shrink the KV cache further.

How does the Mixture of Experts in GLM 4.7 Flash work?+

A router scores 64 experts and activates 4 plus 1 shared, spending just 3B of 30B per token. The Lite routing keeps latency low.

How does GLM 4.7 Flash produce output tokens?+

Final RMSNorm then 154,880 logits. Flash is the cost-optimised GLM-4 tier for high-volume serving.

scene values sourced from:HuggingFace config HuggingFace config

Model ID: glm-4.7-flash

Interactive Architecture

Take GLM 4.7 Flash apart, layer by layer. The 3D teardown below shows its 47 transformer layers (1 dense + 46 MoE), MLA attention, and expert routing, with values from the public model config.

47 layers (1 dense → 46 MoE)

MLA attention · 20Q heads

4+1 of 64 experts active

30B params · 3B active

Token embedding

2048-dim vectors over a 154,880-token vocabulary. The Flash tier shrinks the hidden dim to 2048 for fast, cheap inference.

47 layers (1 dense prefix)

Pre-norm blocks with the first layer dense and the rest MoE, plus a 1-layer MTP head. Roughly half the depth of GLM-4.7.

Multi-head Latent Attention

20 heads over a 512-dim KV latent (q_lora_rank 768). Unlike the GQA flagship, the Flash tier uses MLA-style compression to shrink the KV cache further.

64-expert MoE

A router scores 64 experts and activates 4 plus 1 shared, spending just 3B of 30B per token. The Lite routing keeps latency low.

Output head

Final RMSNorm then 154,880 logits. Flash is the cost-optimised GLM-4 tier for high-volume serving.

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 4.7 Flash use?+

2048-dim vectors over a 154,880-token vocabulary. The Flash tier shrinks the hidden dim to 2048 for fast, cheap inference.

How many transformer layers does GLM 4.7 Flash have?+

Pre-norm blocks with the first layer dense and the rest MoE, plus a 1-layer MTP head. Roughly half the depth of GLM-4.7.

What is Multi-head Latent Attention (MLA) in GLM 4.7 Flash?+

20 heads over a 512-dim KV latent (q_lora_rank 768). Unlike the GQA flagship, the Flash tier uses MLA-style compression to shrink the KV cache further.

How does the Mixture of Experts in GLM 4.7 Flash work?+

A router scores 64 experts and activates 4 plus 1 shared, spending just 3B of 30B per token. The Lite routing keeps latency low.

How does GLM 4.7 Flash produce output tokens?+

Final RMSNorm then 154,880 logits. Flash is the cost-optimised GLM-4 tier for high-volume serving.