XNews

Zhipu AI Releases GLM-5.3 Open-Source Model

Zhipu AI Releases GLM-5.3 Open-Source Model

Zhipu AI released GLM-5.3, an open-source 743B model, on 2026-08-14. - Uses the same base as GLM-5.2; all observable gains attributed to post-training; released hours after a 'sooooooon' post; mixed win/loss with Kimi K3. - Stack: IndexShare for long context, SAO with compaction for long-range RL, slime async framework on Megatron/SGLang; more environments, tasks, and compute. - Programming: internal evaluation shows 50% improvement over GLM-5.2; complex coding and long-range tasks show a generation gap. - Cybersecurity: described as on par with Mythos 5; found 2,436 real vulnerabilities in 269 open-source projects. - Zhipu article: network defense should not be limited to resource-rich institutions; small teams need tools. Hugging Face used GLM-5.2 to defend against OpenAI model intrusion. - Long-range task execution improved for terminal tasks, tool calls, and large codebase repairs.

Key insights

GLM-5.3 Released with Same Base as GLM-5.2

Zhipu AI released GLM-5.3 on 2026-08-14, hours after a 'sooooooon' post by Professor Tang Jie. The model has 743B parameters and uses the same base model as GLM-5.2. All observable capability improvements are attributed to post-training scaling rather than new pre-training. Comprehensive ability is described as mixed win/loss with Kimi K3.

Post-Training Stack Scales Long-Context and RL

GLM-5.3 builds on technical work accumulated on GLM-5.2. The stack includes IndexShare for efficient long-context processing, SAO with compaction for reinforcement-learning gains on long-range tasks, and slime, a large-scale asynchronous post-training framework using Megatron on the training side and SGLang on the rollout side. GLM-5.3 uses more environments, more diverse tasks, and more compute.

Programming Ability Increases 50 Percent Over GLM-5.2

Programming capability is cited as the largest improvement in GLM-5.3. Internal evaluation shows a 50 percent improvement over GLM-5.2. Complex coding and long-range tasks are described as showing a clear generation gap compared with GLM-5.2.

Cybersecurity Capability Rises and Finds Real Vulnerabilities

GLM-5.3 greatly strengthened cybersecurity tasks. Its cybersecurity capability is described as on par with Mythos 5. The model found 2,436 real vulnerabilities across 269 open-source projects.

Open-Source Security Tools for Smaller Teams

Zhipu published an article ahead of the GLM-5.3 open-source release arguing that network defense capability should not be held only by resource-rich institutions. Open-source maintainers, independent developers, and small teams also need such tools. The article says the 5.3 version substantially strengthens cybersecurity tasks.

Security Focus Follows GLM-5.2 Defense Use

GLM-5.3's security optimization followed Hugging Face's use of GLM-5.2 to defend against an OpenAI model intrusion. The release focuses on safety capability and long-range task execution.

Long-Range Task Execution Improves

GLM-5.3 shows large improvements in long-range task execution, including terminal tasks, tool calls, and large codebase repair tasks. The SAO reinforcement-learning algorithm with compaction is designed to keep gains visible on long-range tasks rather than only short tasks.

Source posts