● Covered by 4 sources · 6 reportsMedium impact6 neutral
Z.ai Releases GLM-5.3 with Enhanced Coding and Cybersecurity Capabilities
Key points
- GLM-5.3 released by Z.ai with improved coding and cybersecurity.
- Reportedly found a "serious vulnerability" in Cursor.
- Improvements from scaling post-training, not a new base model.
- Initially available via GLM Coding Plan and ZCode; API and open weights later.
- GLM-5.3 API costs $1.40 per million input tokens and $4.40 per million output tokens.
- Cached input costs $0.26 per million tokens.
- Cached-input storage is free for a limited time.
- GLM-5.3 was released on Friday.
- GLM-5.3 is an updated version of GLM-5.2.
- Post-training optimization included reasoning alignment, supervised fine-tuning, and RLHF.
- LM Studio added GLM-5.3-Flash to its Bionic AI agent platform.
- GLM-5.3-Flash introduces multimodal input and a 1 million-token context window.
- GLM-5.3-Flash is up to 10 times cheaper to run than GLM-5.2.
- GLM-5.3-Flash is a 320-billion Mixture-of-Experts model with 18 billion active parameters.
- GLM-5.3-Flash was anonymously tested as "Ox Alpha" on OpenCode and OpenRouter.
- GLM-5.3 model weights are available on Hugging Face.
- GLM-5.3 license requires security review for companies over $10 billion revenue.
- GLM-5.2 shipped under the permissive MIT license.
- GLM-5.3 improved 50% over GLM-5.2 on Z.ai Code Bench.
- GLM-5.3 achieved open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam.
- GLM-5.3 is state of the art on CyberGym for vulnerability discovery.
- GLM-5.3 more than doubles GLM-5.2 on exploitation benchmarks.
- GLM-5.3 supports deployment with SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth.
- GLM-5.3 supports deployment on Ascend NPU with vLLM-Ascend, xLLM, and SGLang.
- GLM-5.3 supports controlling thinking budget via the reasoning_effort parameter.
GLM-5.3 Enhances Coding and Cybersecurity
Z.ai, a Chinese AI startup, has released GLM-5.3, an update to its GLM series of language models. This new version features substantial gains in long-horizon coding and a notable increase in cybersecurity capabilities. The model's cybersecurity prowess was demonstrated by reportedly identifying a "potentially serious vulnerability in Cursor," an AI coding startup.
Availability and Future Plans
GLM-5.3 is currently accessible through Z.ai's GLM Coding Plan and ZCode coding environment. The company plans to offer API access and open weights approximately two weeks after launch, following safety evaluations and hardening. This phased release strategy indicates a focus on ensuring the model's secure deployment.
Post-Training Scaling Drives Improvements
Unlike previous updates, GLM-5.3 utilizes the same 743-billion-parameter base model as GLM-5.2. All improvements stem from scaling post-training across more diverse environments, tasks, and additional reinforcement-learning compute. This approach tests the extent to which a frontier-scale base model can be advanced without requiring another expensive pretraining cycle, suggesting considerable headroom for existing models.
Unexpected Cybersecurity Gains and Controls
Z.ai noted that cybersecurity capabilities improved faster than anticipated during scaling, particularly in progressing from vulnerability identification to constructing exploitation chains. Due to these advanced capabilities, Z.ai is implementing controls, including a "trusted access" approach for sensitive functionalities, to manage the model's more potent features.
Updates
- GLM-5.3 improved 50% over GLM-5.2 on Z.ai Code Bench.
- GLM-5.3 achieved open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam.
- GLM-5.3 is state of the art on CyberGym for vulnerability discovery.
- GLM-5.3 more than doubles GLM-5.2 on exploitation benchmarks.
- GLM-5.3 supports deployment with SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth.
- GLM-5.3 supports deployment on Ascend NPU with vLLM-Ascend, xLLM, and SGLang.
- GLM-5.3 supports controlling thinking budget via the reasoning_effort parameter.
- GLM-5.3 model weights are available on Hugging Face.
- GLM-5.3 license requires security review for companies over $10 billion revenue.
- GLM-5.2 shipped under the permissive MIT license.
- LM Studio added GLM-5.3-Flash to its Bionic AI agent platform.
- GLM-5.3-Flash introduces multimodal input and a 1 million-token context window.
- GLM-5.3-Flash is up to 10 times cheaper to run than GLM-5.2.
- GLM-5.3-Flash is a 320-billion Mixture-of-Experts model with 18 billion active parameters.
- GLM-5.3-Flash was anonymously tested as "Ox Alpha" on OpenCode and OpenRouter.
- GLM-5.3 was released on Friday.
- GLM-5.3 is an updated version of GLM-5.2.
- Post-training optimization included reasoning alignment, supervised fine-tuning, and RLHF.
- GLM-5.3 API costs $1.40 per million input tokens and $4.40 per million output tokens.
- Cached input costs $0.26 per million tokens.
- Cached-input storage is free for a limited time.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Primary sources
GitHub sgl-project/sglang GitHub vllm-project/vllm GitHub lightseekorg/tokenspeed GitHub huggingface/transformers GitHub kvcache-ai/ktransformers GitHub unslothai/unslothHow outlets covered it
Z.ai has made its GLM-5.3 model weights available on Hugging Face, but introduced a new license that requires companies with over $10 billion in revenue to pass a security review before commercial use. This change impacts large cloud providers and hyperscalers, while individual users retain broad usage rights.
GLM-5.3, an open-weight model, has been released, demonstrating significant improvements in complex coding and long-horizon tasks compared to its predecessor, GLM-5.2. These gains stem from post-training optimizations, making it a notable development for AI applications requiring advanced code generation and cybersecurity functions.
LM Studio has added Z.ai's new GLM-5.3-Flash model to its Bionic AI agent platform, introducing multimodal input, a 1 million-token context window, and reduced pricing compared to GLM-5.2. This integration provides Bionic users with an updated model that offers advanced capabilities and cost efficiency for agentic tasks.
Chinese AI company Z.ai released GLM-5.3, an updated version of its GLM-5.2 model, which shows significant improvements in complex coding and long-horizon tasks through post-training optimization. This release matters as it demonstrates advancements in AI model capabilities for practical engineering and research workflows, potentially impacting developer productivity.
Z.ai has released its GLM-5.3 open-source language model through an API, allowing developers to integrate it into applications. The API pricing remains consistent with the previous GLM-5.2 model, offering a competitive cost structure compared to other high-end frontier models.
Chinese AI startup Z.ai launched GLM-5.3, an updated language model with significant improvements in long-horizon coding and cybersecurity capabilities, which reportedly identified a serious vulnerability in Cursor. The model's advancements come from scaling post-training rather than a new base model, highlighting the potential for existing models to gain new capabilities through further training.