Knowledge Atlas Technology Co., Ltd., branded domestically as Beijing Zhipu Huazhang Technology Co., Ltd. (commonly known as Zhipu AI and rebranded internationally as Z.ai), is a premier Chinese artificial intelligence enterprise specializing in frontier large language models (LLMs), multimodal intelligence, and AI agent architectures. Spun out of Tsinghua University’s computer science department in June 2019, the company was co-founded by leading computer scientists including Chief Scientist Jie Tang, Chairman Debing Liu, and CEO Peng Zhang. Zhipu AI is headquartered in the Haidian District of Beijing, China.
Operating as a standalone pure-play AI pioneer, Zhipu AI achieved a historic milestone on January 8, 2026, by executing its Initial Public Offering (IPO) on the Hong Kong Stock Exchange (HKEX stock code: 02513), establishing itself as China’s first major publicly traded large model developer. Supported by marquee domestic and sovereign backing—alongside global institutional investments from entities such as Prosperity7 Ventures—Zhipu AI has established its General Language Model (GLM) framework as a major technical ecosystem in the global artificial intelligence landscape. Grounded in a "Model-as-a-Service" (MaaS) market paradigm, the enterprise engineers highly capable closed and open-weight models optimized for complex systems engineering, localized multilingual environments, and hardware independence, leveraging native compatibility with domestic chips such as Huawei Ascend, Cambricon Technologies, and Moore Threads.
Core Operations & Technological Infrastructure
Zhipu AI targets the realization of artificial general intelligence (AGI) by integrating data-driven connectionist deep learning with comprehensive, structural knowledge graphs.
1. The GLM Autoregressive Framework
Unlike traditional Western foundational models that primarily rely on left-to-right causal auto-regressive pre-training, Zhipu AI engineered the proprietary General Language Model (GLM) training algorithm. This architecture utilizes an "autoregressive blank infilling" strategy. By randomly removing contiguous segments of input text during pre-training and tasking the network to reconstruct the missing variables, the framework inherently masters bidirectional context capture. This enables the resulting models to excel at simultaneous text comprehension, localized synthesis, context-aware translation, and open-ended generative reasoning.
Operating under restrictive global semiconductor constraints, Zhipu AI concentrates heavy engineering resources on adapting its software layers to alternative compute environments. The enterprise maintains optimized native compatibility protocols that allow its models to execute efficiently across diverse silicon infrastructures:
* Huawei Ascend Deployment: Optimized for cluster-scale pre-training and dense inference workflows on Huawei's Ascend processor lines.
* Cambricon and Moore Threads Integration: Achieving the execution of Mixed-Precision FP8 and Int4 model quantization layers on domestic accelerators, significantly lowering serving overheads and hardware barriers for corporate clients.
3. Agentic Engineering Ecosystems
A primary operational paradigm at Zhipu AI is the transition from static conversational interfaces toward dynamic digital automation. The firm’s ongoing research prioritizes long-horizon planning and execution loops. The platform leverages advanced multi-tool calling capabilities and complex reasoning chains to construct "desktop and mobile automation brains" capable of navigating computing interfaces, interacting with third-party software, and executing human-level workflows autonomously.
Technical Product Lines & Foundation Models
Zhipu AI commercializes its artificial intelligence architectures across three dominant pillars: state-of-the-art base models, developer APIs, and software platforms.
1. The Flagship GLM Large Model Family
The baseline products delivered by Zhipu AI consist of its iterative, multi-parameter foundation models. The models follow a deliberate dual open-source (MIT License) and proprietary commercial distribution route:
GLM-5 Series (Flagship Engine): Introduced in early 2026, GLM-5 represents a massive generational leap, scaling up to 744 billion total parameters structured via a Mixture-of-Experts (MoE) design with 40 billion active parameters per token. Incorporating DeepSeek Sparse Attention (DSA) mechanics to expand context handling while shrinking memory footprints, the model scores near-frontier global parity on developer-centric benchmarks like SWE-bench Verified. It natively acts as an end-to-end autonomous document compiler, converting rough source inputs directly into fully structured production files (.docx, .pdf, .xlsx).
GLM-4.7 & 4.6 Platforms: Highly optimized models engineered for enterprise stability. The series includes highly capable visual-language model (VLM) variants (GLM-4.6V and 4.5V) featuring advanced image processing, spatial token recognition, and real-time visual schema validation.
GLM-4-Voice: An end-to-end native speech large language model engineered to process voice directly without intermediate speech-to-text translation layers. The system allows users to dynamically alter voice outputs, pacing, emotional accentuation, and localized dialects in real time.
2. User-Facing Productivity and Creative Suites
Zhipu Qingyan (BigModel.cn / z.ai): The primary commercial productivity portal. Operating in both a rapid conversational mode and a deeper structured agent mode, the platform provides information processing, coding help, document synthesis, and automated mathematical calculation.
Ying Video Engine: An advanced text-to-video generative AI system that transforms dense alphanumeric strings or baseline reference imagery into cinematic high-definition video clips within seconds, resolving severe visual warping and temporal inconsistency issues.
CodeGeeX Platform: A specialized, low-latency code generation assistant built to integrate directly with native IDE environments, automating refactoring, bug-triaging, and automated test-case scripting workflows.
3. AutoGLM & The OpenClaw System
AutoGLM: A mobile and operating system agent model capable of autonomous multi-step planning. Directed entirely by human voice commands, AutoGLM takes direct control of smartphone interfaces to fulfill long-horizon tasks—such as executing localized purchases, referencing historical shopping trends, or coordinating travel itineraries across separate software applications.
OpenClaw Framework: An open-source personal assistant framework designed to link GLM-5 capabilities natively to personal computers and corporate workstations, enabling physical cross-application device control.
Commercial Services & Enterprise Deployments
Zhipu AI addresses international enterprise, academic, and municipal clients using a highly flexible Model-as-a-Service (MaaS) deployment taxonomy split into two strategic service segments:
1. Cloud-Based Developer Platform (BigModel Open API)
Zhipu AI operates a highly automated public cloud platform providing scalable API tokens to external software developers. The service includes a robust tiered cost model ranging from high-performance paid premium endpoints (GLM-5) to completely free-tier, high-context utility models (GLM-4.7-Flash with 203K tokens), lowering entry overheads for early-stage tech start-ups. The cloud platform natively houses tools for model fine-tuning and API schema validation, alongside real-time search tool integrations.
2. On-Premises Government & Enterprise Customization
For data-sensitive industrial environments, Zhipu AI deploys private edge-computing infrastructures on internal client servers. The team provides custom fine-tuning and retrieval-augmented generation (RAG) structures tailored across the following primary industries:
* Financial Services & Retail Banking: Automated parsing of vast macroeconomic trends, financial compliance reporting, and conversational risk scoring.
* Smart Device Engineering: Embedding scaled-down versions of the GLM architecture natively inside local smart hardware, automotive infotainment centers, and consumer electronics.
* Academic Analytics (AMiner Platform): Leveraging its proprietary AMiner dataset—a multi-million scale academic knowledge graph mapping scholarly publications and global researcher expertise—to deliver advanced data mining, talent tracking, and technological trend profiling tools for universities and sovereign scientific institutions.
Please note: This information is generated by AI. For legal use, verify with company sources.
