Zhipu Unveils GLM-5.3 with 50% Coding Gain, Weights to Open in Two Weeks

Zhipu formally released its new flagship model GLM-5.3 on August 14. The model retains the same base architecture as GLM-5.2 while enhancing capability through an expanded post-training scale. According to the company, internal self-built evaluation showed coding ability improved 50% over the previous generation, and the model achieved first place among open-source models on public benchmarks including Terminal Bench 3.0 and Agents' Last Exam (CLI), with coding and agent capabilities approaching those of Anthropic's top model, Claude Fable 5.

In cybersecurity, GLM-5.3 leads the CyberGym vulnerability discovery test, with the most significant gains in the upstream attack chain. Its performance on attack benchmarks is more than double that of GLM-5.2. The company said it will open the model weights two weeks after release, once safety assessment and reinforcement are completed.

On the same day, JD Cloud integrated GLM-5.3, Zhipu's latest open-source flagship model, and completed adaptation and launch on its model service platform. Separately, DeepSeek released DeepSeek V4 Pro on August 13 and opened testing for its first agent, while US-based SpaceXAI launched Grok 4.6, improving performance in long-process tasks.