GLM-5.2
A flagship model from Zhipu AI made available to all GLM Coding Plan users. It supports extremely long context at the EM level, offers two reasoning intensity modes designated HIGH and MAX, and has an open-source version licensed under MIT.
Core Facts
Timeline (last 90 days)
Official model card scores for Qwen3.8-27B: GPQA Diamond 89.2, LiveCodeBench v6 90.3, SWE-bench Pro 61.7, Terminal Bench 2.1 73.0; Artificial Analysis Intelligence Index score of 52, on par with GPT-5.6 Luna, and just one point below GLM-5.2 (753B) and DeepSeek V4 Pro (1.7T).
Reflection claims Beam's reasoning ability is comparable to GLM-5.2, but its reasoning compute is about 1/3 to 1/4 of the latter's.
Hugging Face suffered an intrusion by an autonomous AI agent in July 2026, and after the security incident it tried to investigate using commercial models but was refused by the provider's security measures, and ultimately ran the open-weight model GLM-5.2 on its own infrastructure to complete the log analysis.
GLM-5.3-Flash scores 63.4 on DeepSWE v1.1, a 17.2-point improvement over GLM-5.2's 46.2, and 55.3 on HLE.
In a blind evaluation organized by Surge AI, professional annotators gave the Mistral Large 4 preview a score of 3.74/5, ranking second among five models, behind Claude Opus 5 (4.22) and ahead of GLM-5.3, Kimi K3, and GLM-5.2.
During a Hugging Face security incident, the Claude Opus and Fable models rejected the majority of vulnerability investigation work due to their guardrail mechanisms, and the team ultimately switched to GLM-5.2 running on their own hardware to continue making progress.
Among model choices, GLM-5.2 is the second-best option with stable overall performance, HY3 is suited for simple webpages, and the Auto default mode is not recommended because it may randomly select models and cause unstable output.
Reflection stated that Beam achieves comparable scores to GLM on reasoning benchmarks while consuming three to four times less inference compute, a figure estimated based on activated parameters and generated token counts rather than measured costs.
Earlier models Claude Opus 4.6 and GLM-5.2 had a success rate of zero on these exploitation tasks.
The team performed one epoch of fine-tuning on 477,642 synthetic reasoning trajectories distilled from GLM-5.2, covering 22,000 curated problems spanning 16 competitive programming contest systems.
32 more timeline events
All Facts (20)
Zhipu released GLM-5.3, claiming its coding capability improved by 50% compared to GLM-5.2.
65%UnverifiedReflection stated that Beam achieves comparable scores to GLM on reasoning benchmarks while consuming three to four times less inference compute, a figure estimated based on activated parameters and generated token counts rather than measured costs.
50%UnverifiedEarlier models Claude Opus 4.6 and GLM-5.2 had a success rate of zero on these exploitation tasks.
50%UnverifiedGLM-5.2 was chosen as the teacher model due to higher accuracy and a roughly 30% reduction in generation length.
50%UnverifiedThe team compared two SFT teacher model approaches based on DeepSeek-V4-Flash and GLM-5.2 respectively, ultimately selecting GLM-5.2.
50%UnverifiedThe team performed one epoch of fine-tuning on 477,642 synthetic reasoning trajectories distilled from GLM-5.2, covering 22,000 curated problems spanning 16 competitive programming contest systems.
50%UnverifiedZhipu did not retrain the model base, but retained GLM-5.2's base and only expanded and reinforced it during the post-training phase.
50%UnverifiedStatements in official announcements such as 'strongest' and 'continuous lead' all come from Zhipu itself and still require corroboration from independent third-party evaluations.
50%UnverifiedThe ModelKey for GLM-5.2 is GLM-5.2.
50%UnverifiedGLM-5.2 continues its positioning as the core engine for Zhipu's ambition to build the strongest domestically produced coding model.
50%UnverifiedGLM-5.2 maintains a continuous lead in the ability to independently complete long-horizon tasks.
50%UnverifiedGLM-5.2 offers an ETA (Estimated Time of Arrival) design that gives users a predictable sense of control.
50%UnverifiedEven the most advanced current models such as GLM-5.2, Claude-Sonnet-4.6, and GPT achieve only just over half marks on this benchmark.
50%UnverifiedThe research team conducted extensive evaluations using multiple agent frameworks on both proprietary models and open-weight models; state-of-the-art models such as GLM-5.2 and Claude-Sonnet-4.6 achieved only slightly above passing performance.
50%UnverifiedState-of-the-art models such as GLM-5.2, Claude-Sonnet-4.6, and GPT-5 achieve only approximately 50–54% Pass@8 on the hard subset.
50%UnverifiedGLM-5.3-Flash has improvements over the previous generation GLM-5.2 in coding capability, robustness of agent interaction, and visual understanding.
50%UnverifiedGLM-5.2 采用 IndexShare 技术，每四个稀疏注意力层复用同一个索引器，在 1M 上下文长度下将每 token 的 FLOPs 降低了 2.9 倍
50%UnverifiedGLM-5.2 首次将长程任务能力建立在稳定可用的 100 万 token 上下文之上
50%UnverifiedGLM-5.2 采用 MIT 开源许可证，官方强调无地域限制、技术访问无国界
50%UnverifiedGLM-5.2 团队大幅扩展了面向编码智能体场景的 1M 上下文训练，覆盖大规模实现、自动化研究、性能优化和复杂调试等任务类型
50%