Model details

GLM-5.3

GLM-5.3 is a post-training-focused release in Z.ai's GLM family, sharing its base model with the earlier GLM-5.2 and building on a stack the team assembled around long-horizon work: IndexShare for efficient long-context processing, SAO for reinforcement learning on long-horizon tasks, and the slime framework for large-scale asynchronous training. All of the gains in this version came from scaling post-training rather than altering the underlying base, using the team's growing set of long-horizon task environments over roughly a month of additional training compute.

The model's practical focus is on complex coding and extended agentic workflows, and Z.ai positions it as the strongest open-weights option for coding work. It shows a large jump over its predecessor on the in-house Z.ai Code Bench, reaches open-source state-of-the-art results on Terminal Bench 3.0 and Agents' Last Exam, and tops the CyberGym benchmark for vulnerability discovery, with the largest gains appearing further along the exploitation chain. These characteristics make GLM-5.3 a good fit for teams that need an open-weights assistant for sustained coding tasks, multi-step problem solving, and security-oriented research where long context and emergent reasoning matter.

Alibaba (China)glm-5.3glm

Open provider page Provider docs

Quick Info

Provider: Alibaba (China)
Model key: glm-5.3
Release date: Aug 14, 2026
Last updated: Aug 14, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $1.10
Output token cost: $3.851

Limits

Output tokens: 131,072 tokens
Context window: 1,000,000 tokens

Transparent token rates

Compare GLM-5.3 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about GLM-5.3

Alibaba (China)

CoveragePreview

GLM-5.3 marks a notable shift away from the permissive licensing of the GLM-5 line: whereas GLM-5.2 and earlier GLM-5 releases shipped under MIT, GLM-5.3 is released under a custom license with a revenue-gated clause for inference and fine-tuning providers. The clause states that if a Licensee (or any of its affiliates The same piece situates this licensing change within a broader 2026 pattern in which Chinese frontier model makers (citing Kimi K3 and MiniMax M3) have moved toward more restrictive, revenue-threshold and use-case-prohibited licenses, while Western model makers such as Google and Meta have moved toward Apache 2.0. For

Alibaba (China)

Coverage

Z.ai announced its open-weight GLM 5.3 model on August 14, 2026, describing it as built for advanced coding and cybersecurity work on the same base architecture as GLM 5.2, with all capability gains attributed to expanded post-training rather than a larger base. According to the report, GLM 5.3 reached 84.5% on the Cyb Alongside the model, Z.ai launched OpenVuln (branded "VulnHunter"), a scanning service that uses GLM 5.3 to check public code repositories and publish an aggregate security score while keeping detailed findings private until fixes are ready. Z.ai's disclosure ledger credits GLM 5.3 with 2,436 vulnerability findings acr

Alibaba (China)

Coverage

China Daily's launch-day report confirms that Zhipu AI (also known as Z.ai) launched GLM-5.3 on August 14, 2026, as its next-generation foundation model claiming coding and AI agent capabilities approaching "Claude Fable 5." The report states GLM-5.3 has 743 billion parameters, with most performance improvement coming According to the Zhipu technical report cited by China Daily, GLM-5.3 scored 28.3 on Terminal-Bench 3.0 (claimed highest among open-source models and ahead of Moonshot AI's Kimi K3) and 66.9 on DeepSWE 1.1, another software engineering benchmark. Zhipu said GLM-5.3 ranked as the strongest open-source model across sever

Alibaba (China)

CoverageBenchmark

GLM-5.3, announced on August 14, 2026, is a post-training-only update that Z.ai explicitly states uses the same base model as GLM-5.2, with all gains attributed to scaled post-training. Per Z.ai documentation summarized in the page, GLM-5.3 is text-only, always-reasoning, with a 1M-token context window, a 128K max outp The same page situates GLM-5.3 within the broader GLM-5 family: GLM-5.2 (weight repo created on Hugging Face on 2026-06-16) shipped with MIT-licensed open weights, a 1M-token context, an IndexShare architecture change, an improved MTP layer for speculative decoding, and Z.ai's positioning as the highest-ranked open-sou