Z.ai, the Beijing artificial-intelligence developer co-founded by Zhang Peng, Tang Jie and Liu Debing, published an August 14 research post presenting GLM-5.3 as a post-training update for coding agents and vulnerability research. The GLM-5.3 post says the model uses the same base model as GLM-5.2, with gains coming from another month of reinforcement learning, expanded task environments and additional post-training compute.
The product record is less clear. As of August 14, Z.ai's official release notes identified GLM-5.1 as the latest documented GLM-5 release and did not list GLM-5.3. Z.ai's GLM-5 documentation describes GLM-5 as an open-weight foundation model available through the GLM Coding Plan, without identifying a GLM-5.3 API endpoint, hosted-model identifier or downloadable artifact.
Z.ai says in the August 14 post that it will release GLM-5.3 weights two weeks after launch, once safety evaluation and hardening are complete. The post links readers to Z.ai's Coding Plan and ZCode, Z.ai's coding agent, while labeling Hugging Face access "Coming Soon." It does not publish an API endpoint, model identifier or weight download. Developers therefore lacked a confirmed GLM-5.3 access path or reproducible production configuration as of August 14.
The August 14 page is a Z.ai-authored research post containing detailed training and benchmark claims. Its scores and cyber findings should be treated as Z.ai's claims until independent evaluations reproduce them. The gap between the post and Z.ai's product documentation also leaves GLM-5.3's eventual license and serving configuration unconfirmed.
Tang, a Tsinghua University computer science professor and Z.ai co-founder, created AMiner, an academic search and knowledge-graph system that has operated since 2006. His research has covered data mining, social networks, knowledge graphs and foundation models, and his group contributed to GLM-130B, ChatGLM, CogView, CogVideo and CodeGeeX. His Tsinghua biography says he earned his doctorate in computer science there in 2006. On his research homepage, Tang describes his mission as "teaching machines to think like humans."
Zhang, Z.ai's CEO, worked with Tsinghua's Knowledge Engineering Group and its data-science research center before helping commercialize that research. Liu, Z.ai's chairman, has focused on corporate and capital operations. The founders built Z.ai, previously known outside China as Zhipu AI, around research originating at Tsinghua.
Z.ai has substantial resources behind its model work. Its Hong Kong offering prospectus reported 883 employees as of June 30, 2025, including 657 in research and development. Revenue reached RMB312.4 million in 2024 and RMB190.9 million during the first six months of 2025. At midyear, Z.ai said it served more than 8,000 institutional customers and approximately 80 million devices.
The same prospectus says Z.ai had more than 12,000 institutional customers for the nine months ended September 30, 2025, and recorded average daily token consumption of approximately 4.2 trillion in November 2025. Those are Z.ai-reported measures disclosed for the listing process, rather than independently audited indicators of active usage.
The prospectus reported approximately RMB8.364 billion in pre-IPO investments across rounds from Series Angel through Series B6. The latest listed round brought in RMB4.377 billion and implied a post-money valuation of approximately RMB24.377 billion. Named backers disclosed in the prospectus include Meituan, Qiming Venture Partners, Legend Capital, Ant Group and Aramco Ventures. The global offering covered 37,419,500 shares at HK$116.20 each, implying gross proceeds of roughly $557 million.
CB Insights reports that Z.ai raised more than $1.4 billion in private capital. Other reported backers included Alibaba, Tencent, Hillhouse and Chinese government funds, according to Forbes. Z.ai listed in Hong Kong on January 8, 2026, in an offering that Dealroom reported raised about $558 million.
Post-training drives the reported gains
Z.ai says the GLM-5.3 training stack uses three components introduced with GLM-5.2: IndexShare for efficient long-context processing, SAO for reinforcement learning on long-running tasks, and slime, its open-source framework for asynchronous reinforcement-learning training.
The training environments described in the post resemble units of professional engineering work instead of short coding exercises. In one example, a model receives access to compute clusters, storage, documentation, code and experiment results. It must locate a machine-learning infrastructure bottleneck, change the implementation, run tests and produce a measurable speed improvement without breaking correctness.
Building enough of those environments has become its own engineering problem. Z.ai says research agents collect task patterns and turn them into executable scenarios with hidden state and multi-step dependencies. A separate judge attempts each task to establish that it can be solved, while generated verifiers test the result and try to close shortcuts that could let the model earn a reward without completing the underlying work.
Humans remain part of that pipeline. Z.ai acknowledges that environment creation and verification still require meaningful manual involvement. The approach will be difficult to scale unless Z.ai can produce reliable tasks and rewards faster than specialists can design and audit them by hand.
Z.ai's coding benchmark table reports a rise from 4.6 to 28.3 on Terminal-Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1 and from 23.8 to 28.5 on Agents' Last Exam. On its private Code Bench, Z.ai claims a 50% improvement. At the maximum effort setting, Z.ai says GLM-5.3 completed 34.5% of tasks using roughly 75,000 output tokens per task, compared with 23.4% at 96,000 tokens for GLM-5.2. The private test's task set and methodology cannot be independently inspected through the post, and none of these results has been independently reproduced in the supplied material.
The comparison table does not produce a single leader. It places GLM-5.3 against models from Moonshot AI, DeepSeek, Alibaba, Anthropic and OpenAI. GLM-5.3 leads the listed open-weight alternatives on some evaluations, while competitors post higher results on several coding and exploitation tests. Z.ai's description of GLM-5.3 as the strongest open-weights coding model depends on its selected evaluations and settings.
Z.ai reports an unexpected cyber gain
The more consequential claim appears outside the coding pitch. Z.ai says GLM-5.3's ability to find and exploit software vulnerabilities improved faster than expected as post-training scaled. The post describes a progression from identifying isolated flaws toward forming multi-stage exploitation plans.
In its cyber benchmark results, Z.ai reports a CyberGym score of 84.5 for GLM-5.3, up from 77.2 for GLM-5.2. It lists a 54.4 ExploitBench score, compared with 24.4 for GLM-5.2. On ExploitGym, Z.ai reports 105 and 130 successful results at two-hour and six-hour limits, compared with 29 and 39 for GLM-5.2. These results require independent reproduction.
The comparison contains an internal naming discrepancy. Z.ai's table labels Anthropic's 78.0 ExploitBench result "Fable 5," while the surrounding text calls the model "Mythos 5." The post does not establish which configuration Z.ai tested, so the result cannot reliably be assigned to either name. The same table lists GPT-5.6 Sol at 76.5.
The practical meaning of the cyber figures depends on each benchmark's denominator, tool access, retry policy and execution conditions. Raw exploit counts are especially difficult to compare without that protocol detail. Z.ai's evaluation stack nevertheless records its largest relative gains farther along the path from vulnerability discovery to working exploitation.
Beyond controlled benchmarks, Z.ai reports that its models identified 2,436 vulnerabilities across 269 projects after expert review, screening and deduplication, including 1,097 findings described as critical or high severity. Z.ai says 53 findings were publicly disclosed and 2,383 remained under embargo, with updates recorded in the Z.ai Security Disclosure Ledger.
Z.ai attributes the findings to work with several security teams in China. The listed targets include system kernels, operating systems, browser engines, open-source infrastructure, web applications and network protocols. These remain Z.ai-reported totals rather than an independently audited measurement, and individual findings remain subject to expert review and coordinated disclosure.
Open weights and coding distribution are separate contests
Z.ai competes with DeepSeek, Moonshot AI's Kimi and Alibaba's Qwen in open-weight models. Its commercial distribution also depends on coding products that put models directly into developer workflows. ZCode and Z.ai's Coding Plan provide that product layer, while downstream tools such as Cursor compete by controlling the integrated coding environment instead of supplying only a foundation model.
On its subscription page, Z.ai advertises pricing starting at $16.2 per month for its Lite coding plan and says the plan supports more than 20 coding tools. Hosted subscriptions give Z.ai a direct billing relationship with developers. Open-weight deployment serves buyers that want to run and modify a model on their own infrastructure. Z.ai's official materials describe GLM-5 as open weight, while the GLM-5.3 post promises weights only after a two-week safety evaluation and hardening period.
The model described in the research post is intended for jobs that can consume tens of thousands of tokens and operate across terminals, repositories and infrastructure. Better benchmark scores could help Z.ai sell that use case. Adoption will depend on reliable completion of long-running work, reproducible evaluations and a documented route to the model.
The reported cyber results expose a safety cost in the same training method. An agent trained to inspect unfamiliar code, operate tools, test hypotheses and persist through failures acquires mechanics that can support offensive security work. Z.ai says its researchers encountered that capability growth while pursuing stronger coding performance.
For now, Z.ai has published detailed GLM-5.3 training, coding and cyber claims without adding GLM-5.3 to the official release notes. The research post promises weights after a two-week review, while Z.ai's documentation provides no GLM-5.3 API endpoint, hosted-model identifier or downloadable artifact as of August 14. Independent testing cannot resolve the benchmark claims until researchers can establish which artifact or serving configuration corresponds to the reported results.