Z.ai

Z.ai matters as a bridge between Tsinghua's academic model research, China's commercial foundation-model market, and the international open-weight ecosystem. Its GLM releases provide technically substantial Chinese and English alternatives for coding and agent work. The main evidence caveat is that release speed and company benchmark disclosure currently outpace independent reproduction, especially for the newest cyber and frontier-comparison claims.^1,4,6,8

Official site ↗Models in the Atlas →

Profile

Overview

From Tsinghua research to Zhipu AI

Z.ai is the international brand of Beijing Zhipu Huazhang Technology, usually known in China as Zhipu AI. The company was founded in 2019 to commercialize work from Tsinghua University's Knowledge Engineering Group. Its official history traces a path from knowledge-graph research and the AMiner academic platform to the GLM model family. Tang Jie and Li Juanzi are associated with the academic origin, while Zhang Peng has led commercialization as chief executive. The Z.ai name became the global-facing brand in 2025, but Zhipu AI remains common in Chinese and financial reporting.^1,2

The GLM research line

The research lineage predates the current chatbot market. The 2021 GLM paper proposed a pretraining framework based on autoregressive blank infilling that combined bidirectional context with autoregressive generation. GLM-130B then scaled the approach into an open bilingual English and Chinese model and documented training stability techniques. ChatGLM adapted the family for dialogue, while later releases shortened the public name to GLM and expanded into vision, coding, tool use, and long-context agent tasks.^3,4

Open models and specialized systems

The organization has repeatedly paired general models with specialized open artifacts. CodeGeeX provided a multilingual code-generation model and HumanEval-X, a benchmark spanning several programming languages. GLM-130B published a large bilingual checkpoint and technical report. Recent GLM releases use mixture-of-experts architectures and MIT-licensed weights for some checkpoints, while the company also sells subscriptions and API access. Availability varies by model, so open weights should not be read as a promise that every frontier checkpoint ships immediately.^5,4,1

A rapid 2026 release cycle

In 2026 the company accelerated from GLM-5 through GLM-5.3 and GLM-5.3-Flash. Z.ai emphasized coding, long-horizon agents, one-million-token context, and lower serving cost. It delayed the GLM-5.3 weight release after reporting strong cyber capabilities, then published the weights on August 28 after the announced two-week review window. Axios independently reported the delay and the associated dual-use concern, but the exact benchmark comparisons still originate largely with Z.ai. The current factual record supports rapid technical progress; broader claims of parity with closed US frontier labs remain less certain until more independent evaluations accumulate.^6,7,8,9,10

Company evidence

Epoch AI dataset ↗

Reported revenue: $153MPeriod interpolation; Oct 1, 2025 · Likely[1][2]
Latest funding: Not disclosed$39.4B post-money valuation; Mar 31, 2026 · Confident[1]
Reported staff: 883Full company; Jun 30, 2025 · Confident[1]

Reported estimates, not audited figures. Confidence labels and source links are preserved from the dataset.

Reporting and context

Notable contributions

01General Language Model pretrainingThe GLM research program developed autoregressive blank-infilling objectives that combine bidirectional context encoding with autoregressive generation in one pretrained model.³
02Open bilingual scaling with GLM-130BGLM-130B released a 130-billion-parameter English and Chinese model with a technical account of architecture, training stability, scaling behavior, and evaluation.⁴
03CodeGeeX and HumanEval-XCodeGeeX paired an open multilingual code model with HumanEval-X, extending code-generation evaluation across several programming languages rather than relying on Python alone.⁵
04Open agent-oriented GLM releasesRecent GLM checkpoints package long context, coding, tool use, and reasoning for agent workflows while publishing weights for independent deployment, subject to model-specific release timing and licenses.^6,7,8

Sources · 10+

Model lab #7

As of Aug 29, 2026

GLM-5.3's weights have landed; independent validation is now the next test.

GLM-5.3 and GLM-5.3-Flash extend Z.ai's fast 2026 cadence across coding, long-context agent work, and lower-cost inference. Z.ai released the GLM-5.3 weights on August 28 after its announced two-week safety-review window, completing the staged release rather than leaving the model under hold. The pause remains a meaningful governance signal, but the more important open question is whether the additional review materially reduced risk. Confidence remains medium because many frontier comparisons still depend on Z.ai's own evaluations and independent reproduction is limited.

Why this editorial standing+

Z.ai keeps the seed score because its recent model evidence and release cadence are unusually strong: GLM-5.2, GLM-5.3, and GLM-5.3-Flash extend the family across long-context agents, coding, and efficient inference. The score remains below the category leaders because much of the freshest performance evidence is self-reported, its international ecosystem is smaller, and independent scrutiny of governance and deployment remains limited.^1,6,7,8,10

Model evidence: 33 / 35
Research influence: 20 / 25
Release momentum: 20 / 20
Ecosystem: 8 / 10
Trajectory: 8 / 10

Notable people

ZP
Zhang PengChief executivecurrent
TJ
Tang JieCo-founder and chief AI officercurrent
LJ
Li JuanziCo-founder and Tsinghua research collaboratorcurrent

Revision history1 updates +

Initial deep research packet
Aug 29, 2026
Expanded the seed with the Tsinghua and Zhipu history, GLM papers, CodeGeeX, current releases, and explicit uncertainty around self-reported frontier comparisons.