GLM-5.3

Z.ai announced GLM 5.3 on August 14, 2026, positioning it as an open-weight model built for advanced coding and cybersecurity work that shares the same base architecture as GLM 5.2, with every capability gain attributed to expanded post-training rather than a larger base model. The release launched alongside OpenVuln ( On Z.ai's self-reported benchmarks, GLM 5.3 reached 84.5% on CyberGym (up from 77.2% for GLM 5.2) and 54.4% on ExploitBench (more than double GLM 5.2's 24.4%), though the article notes these figures have not been independently verified. Citing the dual-use risk of those same gains for attackers, Z.ai said it would hold

GLM-5.3 was released on August 14, 2026 via the GLM Coding Plan with 743B parameters, reusing the identical GLM-5.2 base model. Z.ai improved coding and agentic capability solely through post-training, including reinforcement learning and agent-oriented fine-tuning, without modifying the pretraining corpus or architecture. As of the article, weights were not yet public, with release planned around late August 2026 pending security review. Z.ai's internal evaluations claim roughly a 50% improvement in coding ability over GLM-5.2, and the model posts top scores among open models on Terminal-Bench 3.0 (28.3%), Agent's Last Exam (28.5%), and AutomationBench. Since the base model is unchanged from GLM-5.2, the general scale of compute needed for inference carries over from that release.

Z.ai released GLM-5.3 on 14 August 2026, according to a launch-partner technical brief published the same day by Qubrid AI. GLM-5.3 uses the same 743B-parameter base as GLM-5.2, with all reported capability gains coming from scaled post-training (more executable environments, more environment types, and longer RL runs) GLM-5.3 ships with three reasoning effort levels (low, high, max, with max as default) and disabling thinking is no longer supported. Reported benchmark jumps include Terminal-Bench 3.0 rising from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, CyberGym at 84.5% (top of Z.ai's launch chart), and GDPval-AA v2 climbing fro

Z.AI (formerly Zhipu AI) released GLM-5.3 on August 14, 2026, claiming it is the highest-ranked open-source model across multiple mainstream benchmarks, with coding and agent capabilities approaching Anthropic's Claude Opus 5. GLM-5.3 retains the same base model as its predecessor GLM-5.2, but extensive post-training scaling has significantly raised its intelligence ceiling. GLM-5.3 scored 28.3% on Terminal-Bench 3.0, up from 4.6% for its predecessor, compared with 42.7% scored by Anthropic's Claude Opus 5. On DeepSWE v1.1, focused on long-horizon software engineering, the model jumped to 66.9% from 46.2%, versus Claude Opus 5's leading 74%. Other Chinese open models mentioned include Moonshot AI's Kimi K3 and DeepSeek's V4 Pro.

Zhipu AI (Z.ai) released GLM-5.3 on August 14, 2026, describing it as the strongest open-weight model for coding. The Beijing-based company said it plans to publish the model's weights roughly two weeks after launch, following additional security testing. GLM-5.3 retains the same base model as GLM-5.2, with all gains coming from post-training rather than a new architecture. According to Z.ai, GLM-5.3 delivers a 50% improvement in coding performance over GLM-5.2 in internal evaluations and ranks first among open-weight systems on Terminal Bench 3.0 and Agents' Last Exam (CLI). The company attributed the gains to expanded task environments and its IndexShare, SAO, and next-generation Slime reinforcement-learning frameworks, and noted an unexpected improvement in cybersecurity-related tasks.

GLM-5.3, announced on August 14, 2026, is a post-training-only release using the same base model as GLM-5.2, with every gain coming from post-training. Z.ai frames it as Frontier Coding with Emergent Cyber Capabilities: a 50% coding gain over GLM-5.2 on the in-house Z.ai Code Bench, claimed open-source SOTA on Terminal-Bench 3.0 and Agents' Last Exam (CLI), and unexpectedly fast growth in vulnerability discovery and exploitation ability. Working with security teams in China, Z.ai reports the model surfaced 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues, with an average latent lifetime of 26.6 years and the oldest flaw introduced in 1981. GLM-5.3 is text-only, always reasoning, with a 1M-token context, 128K max output, and three effort levels (low, high, max, default max); disabling thinking is no longer supported.