Z.ai

Z.ai is a Chinese AI company formerly known as Zhipu, developing the GLM family of language models and services for developers and enterprises. Its GLM-4.5 generation brought reasoning, coding and agentic abilities together; subsequent GLM-5 models expanded model scale, sparse attention and reinforcement learning for complex engineering tasks.[4][6]

Those agent-oriented models followed several years of work on pre-training frameworks, bilingual language models and locally deployable conversational models.[5][4]

Founded in 2019, the company developed the GLM pre-training framework at the end of 2020 and introduced GLM in March 2021. It trained the 10-billion-parameter GLM-10B in 2021, alongside a trillion-parameter sparse model using a mixture-of-experts architecture.[5][4]

In September 2022, it open-sourced GLM-130B, a 130-billion-parameter model pre-trained in English and Chinese. ChatGLM followed in March 2023; the open-source ChatGLM-6B enabled developers to fine-tune and deploy a conversational model locally. The company’s historical timeline records GLM-4 in January 2024, AutoGLM in October and GLM-Zero-Preview in December.[5]

Z.ai released GLM-4.5 on July 28, 2025, with 355 billion total parameters, alongside the 106-billion-parameter GLM-4.5-Air. Both integrated reasoning, coding and agentic abilities in a single model. The company described multi-step planning and end-to-end workflows as intended uses and announced an open, auditable license with options for on-premise deployment and fine-tuning.[4]

By that July announcement, its product suite included Z.ai, QingYan, CodeGeeX, CogVLM and CogView, plus a Model-as-a-Service platform for developers and enterprises. Z.ai reported more than 40 million cumulative worldwide downloads of its open models—a portfolio download count rather than a count of distinct users or paying customers.[4]

GLM-5 launched in February 2026, targeting complex systems engineering and long-horizon agentic tasks. Compared with GLM-4.5, it increased total parameters from 355 billion to 744 billion, active parameters from 32 billion to 40 billion, and pre-training data from 23 trillion to 28.5 trillion tokens. It also incorporated DeepSeek Sparse Attention, which Z.ai said reduced deployment costs while preserving long-context capacity.[1][6]

For post-training, Z.ai developed slime, asynchronous reinforcement-learning infrastructure designed to improve training throughput and support finer-grained training iterations. The GLM-5 repository credits the technical report to GLM-5-Team and a large coauthor group, including Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen and Jie Tang.[6]

One concrete test of GLM-5’s intended planning abilities was Vending Bench 2, which asks a model to operate a simulated vending-machine business over a one-year horizon. Z.ai reported a final account balance of $4,432. Deployment brought a separate constraint: Latent.Space’s February 12 account reported that Zhipu warned of limited serving capacity, delaying broader Coding Plan access and prompting pricing changes.[6][1]

GLM-5.2, covered in June 2026, extended the family to a one-million-token context. Its IndexShare mechanism reuses an indexer across every four sparse-attention layers; Z.ai reported a 2.9× reduction in per-token FLOPs at that context length. This is a context-specific computation claim, rather than a universal serving-speed improvement.[6][3]

Practitioner reception also became part of GLM-5.2’s story. Latent.Space’s June 19 account relayed Jeremy Howard’s assessment that it was “at least as good as Opus 4.8 and GPT 5.5” for his use, while noting its lack of vision support.[3]

With GLM-5.3, Z.ai retained GLM-5.2’s base model and attributed the gains to post-training. It reported a 50% improvement on its in-house Z.ai Code Bench. In commentary relayed by Latent.Space on August 20, 2026, Jie Tang argued that parameter count should be considered alongside data, compute allocation and the conditions under which a model will run.[7][2]

The August account described training environments modeled on substantial professional work. In an ML infrastructure task, an agent could access compute clusters, storage, documentation, code and experiment results, then diagnose bottlenecks, implement optimizations and demonstrate a speedup while preserving correctness. To scale such training, the team described pipelines that synthesized runnable environments and, for some tasks, reward signals. Research agents assembled tasks with multi-step dependencies and hidden state; a judge agent checked solvability. Verifiers were synthesized without access to reference solutions, while solver trajectories helped identify reward shortcuts.[2]