NUWA 来源档案
GLM-5.2 — official Z.ai Hugging Face model card
完整核查了 GLM-5.2 官方模型卡及评测方法。模型卡披露的是通用能力与使用信息，未披露专门安全评测或前沿风险结论；编码和工具使用分数不作为直接风险结果。
范围与归属
- 数据报告方
- Z.ai
- 实际评测方
- 各观测的执行方未完整确定
- 抓取版本
- 2026-10-26
- 来源身份
- 75f1d77e20e78286
来源转录不等于实验复现。未确定的日期、执行方及配置保留为未知；原文分组不作为经核验排名。
来源发现
以下陈述整理自原始来源，未作为 NUWA 独立实验结论。
这份完整模型卡主要介绍能力、使用方式和基准评测方法，未披露专门的安全基准结果或前沿风险结论。
README.md:1-108结论范围限于原始来源的任务与实验条件；原文定位保留供核查。
开发者将 GLM-5.2 定位为面向长时程任务、支持更大上下文且可灵活调节编程推理投入的模型。
README.md:31-37结论范围限于原始来源的任务与实验条件；原文定位保留供核查。
NL2Repo 的评测说明提到，使用规则检查和大模型判定来防范针对基准评测的投机行为。
README.md:83结论范围限于原始来源的任务与实验条件；原文定位保留供核查。
DeepSWE 和 ProgramBench 的方法说明采用了禁用互联网访问的隔离执行环境；各软件基准的资源与时间预算有所不同。
README.md:84-87结论范围限于原始来源的任务与实验条件；原文定位保留供核查。
模型卡将 FrontierSWE 的评测归于 Proximal，将 PostTrainBench 的评测归于 PostTrainBench，将 SWE-Marathon 的评测归于 Abundant AI。
README.md:90-92结论范围限于原始来源的任务与实验条件；原文定位保留供核查。
官方渲染页面另有一条 Agents Last Exam 评测记录，其说明明确称该记录来自排行榜导出，并非经过 Hugging Face 验证的评测。
model-card-page.txt:639-645结论范围限于原始来源的任务与实验条件；原文定位保留供核查。
提取范围
以下条目说明来源各部分的处理状态。个别观测已核验，不代表整份文档的全部结论和结果均已验证。
查看 16 项覆盖记录
| 章节 | 处理状态 | 定位与范围 |
|---|---|---|
| Card metadata / source identity | extracted | README.md:1-29; model-api.json $.id, $.author, $.sha, $.createdAt, $.lastModifiedOfficial zai-org repository. API SHA and raw response X-Repo-Commit agree. Repository creation/modification timestamps are not asserted as card publication/revision dates. |
| Introduction | qualitative | README.md:31-39Developer describes long-horizon capability, context capacity, coding and open licensing. These are not safety evaluation results. Architecture efficiency numbers are outside this bounded risk extraction. |
| Benchmark: Reasoning | not-relevant | README.md:45-53Generic reasoning/math/science scores; no dedicated dangerous-capability or safety evaluation is identified. No subject-task content or numerical biology comparisons extracted. |
| Benchmark: Coding | extracted | README.md:54-63;82-87;90-92Exact GLM-5.2 entries retained only as general software/long-horizon capability context; no inference to cyber offense, self-improvement, loss of control, or risk thresholds. Comparative model columns omitted. |
| Benchmark: Agentic | extracted | README.md:64-66;88-89Exact GLM-5.2 tool-use scores retained as capability context, not an autonomy risk result. |
| Serve GLM-5.2 Locally | not-relevant | README.md:68-77Framework compatibility / deployment links do not disclose safety testing; no setup commands executed. |
| Footnote | extracted | README.md:79-92Relevant harnesses, budgets, trial count, external evaluator attribution, public task count and high-level evaluation controls recorded. The NL2Repo anti-hacking measure has no measured failure rate. |
| Citation | qualitative | README.md:94-108References GLM-5 technical report (2602.15763); it is not a GLM-5.2 safety report. Linked sources are dependencies, not extracted or inherited evidence. |
| Hugging Face Evaluation results widget | not-relevant | model-card-page.txt:614-648; model-card-page.html Evaluation resultsPage includes platform evaluation entries including an Agents Last Exam leaderboard export expressly not HF-verified. Those are outside developer README and are not treated as developer risk results; duplicates of card scores are not added. |
| Dedicated safety evaluation | not-disclosed | Entire pinned README.md:1-108; official page model-card bodyNo safety benchmark, refusal/harmlessness result, jailbreak rate, or adversarial safety result is disclosed in this card. |
| Frontier risk policy / thresholds / deployment decision | not-disclosed | Entire pinned README.md:1-108; official page model-card bodyNo frontier-risk taxonomy, capability threshold, risk tier, safety case or deployment approval conclusion is disclosed. |
| CBRN / biological / chemical / nuclear risk | not-disclosed | Entire pinned README.md:1-108; official page model-card bodyNo dedicated CBRN dangerous-capability assessment, uplift study, or mitigation evidence is disclosed. No dangerous task content or comparisons extracted. |
| Cybersecurity dangerous capability | not-disclosed | Entire pinned README.md:1-108; official page model-card bodyNo offensive cyber capability or vulnerability exploitation safety assessment is disclosed. Generic coding results do not establish cyber-risk capability. |
| Autonomy / self-improvement / control | not-disclosed | Entire pinned README.md:1-108; official page model-card bodyNo escape, replication, persistence, resource acquisition, autonomous R&D risk threshold, or loss-of-control assessment is disclosed. |
| Deception / scheming / manipulation / oversight | not-disclosed | Entire pinned README.md:1-108; official page model-card bodyNo deceptive alignment, scheming, persuasion risk, sandbagging, oversight integrity, or situational-awareness evaluation is disclosed. |
| Product safeguards and monitoring | not-disclosed | Entire pinned README.md:1-108; official page model-card bodyNo measured product safeguards, red-team safety outcomes, post-release incident monitoring or residual-risk assessment is disclosed. Benchmark sandboxing and anti-hacking controls are narrower evaluation methods. |
来源待核实事项
This official model card discloses capability benchmarks and usage information but no dedicated safety evaluation, frontier-risk classification or deployment safety conclusion. This disclosure gap is limited to the reviewed card.
Software benchmark sandboxing and anti-hacking checks describe evaluation controls without measured product safeguard effectiveness. Externally conducted benchmark entries and a platform leaderboard widget retain their separate provenance.
版本指纹
https://huggingface.co/zai-org/GLM-5.2/raw/cf457fa734ab149ffef225f80893eb38c6ff5cdc/README.md
7ba22185651d8f69fbf6227ee718eb680fd834ec8e79ba78df6fa54896614b3dhttps://huggingface.co/zai-org/GLM-5.2
fce38778229b2fe8d443d5af4cb0ae56ca644ada69b05f5f2b373009f5328b80https://huggingface.co/api/models/zai-org/GLM-5.2
064e59e4acce4f697ea7f67ebd8aeff954cba9598ea0aaef5708429a629356e4https://huggingface.co/zai-org/GLM-5.2
73eb98c1a6282cddb11fc787cf915d42310da6ab164807ca4fbf6dd8cd0ed932