摘要
智谱发布 GLM-5.3，沿用 GLM-5.2 的同一基座，提升全部来自后训练（官方称「scaling post-training 是我们唯一做的事」）。官方将其定位为最强开源权重编程模型，主攻网络安全和 agent 能力：与中国安全团队合作，在 269 个项目中发现 2,436 个漏洞，并在 cvd.z.ai 公开登记。模型现已通过 GLM Coding Plan 提供，兼容 Claude Code 和 OpenCode。API 即将上线，开源权重计划在约 2 周的安全审查之后放出。
为什么重要
对做 coding agent 的人来说，这是一个有价值的数据点：只靠后训练就号称拿下开源权重编程 SOTA，并在 CyberGym 上超过美国前沿模型，说明后训练扩展仍能显著提升 agent 和编程能力。「安全审查通过后再放权重」这种做法本身也值得跟踪。
技术细节
| 上下文 | 1M tokens，最大输出 128K |
|---|---|
| Z.Ai Code Bench | 对比 GLM-5.2 +50% |
| Terminal-Bench 3.0 | 28.3（此前 4.6） |
| DeepSWE V1.1 | 66.9（此前 46.2） |
| Agent'S Last Exam | 28.5（此前 23.8） |
| CyberGym | 84.5%（对比 Mythos 5 83.8%、GPT-5.6 Sol 83.6%） |
| ExploitBench | 54.4%（此前 24.4%） |
| 发现漏洞 | 覆盖 269 个项目共 2,436 个（注册库：cvd.z.ai） |
| 权重 | 2026-08-25 落地：zai-org/GLM-5.3（FP8，153 个文件）+ GLM-5.3-BF16，仓库 2026-08-25T06:42Z 创建（较 ~8/28 预估提前约 3 天）；无门禁；总参数 753B |
| 许可证 | 自定义 glm-5.3 许可证：MIT 式；唯一限制——年收入超 100 亿美元的 Model-as-a-Service 业务商用前需通过 Z.ai 安全审查 |
| 独立评测 | Artificial Analysis Intelligence Index 60（GLM-5.3 max；与 Kimi K3 持平、较 GLM-5.2 +7）；GLM-5.3-Flash 57（8/26 上架） |
后续更新
2026-08-15 Z.ai published follow-up 'Preparing GLM-5.3 for Open Release: A Responsible Path to Cyber Defense' at 2026-08-15T06:32Z, plus live OpenVuln demo space (huggingface.co/spaces/zai-org/OpenVuln). Reaffirms weights-after-review plan.
2026-08-25 Recovered from index (previous run wrote the update note to index/events.json only): open weights landed: HF zai-org/GLM-5.3 (FP8, 153 files) + GLM-5.3-BF16 created 2026-08-25T06:42Z (~3 days ahead of the ~8/28 estimate); ungated; 753B total params; custom glm-5.3 license (MIT-style; only restriction: >$10B-revenue MaaS businesses need Z.ai security review before commercial use); independent eval: Artificial Analysis Intelligence Index 60 (max) — on par with Kimi K3, +7 vs GLM-5.2; captured 2026-08-28 run; recommendation stays TRIAL
2026-08-29 Model card on HF fully rewritten (repo 'Initial commit 0828' at 2026-08-27T17:16:16Z, coverage-gap recovery): complete benchmark table with per-benchmark footnotes is now public — states GLM-5.3 uses the same base model as GLM-5.2 (all gains from post-training); Terminal Bench 2.1 88.2 / TB 3.0 28.3 / DeepSWE v1.1 66.9 / CyberGym 84.5 / ExploitGym 105@2h / HLE w-tools 62.5 / GDPval-AA v2 1769; FrontierSWE 78.1 evaluated by Proximal; most evals run in a Claude Code 2.1.207 harness. Community PRs #2 (2026-08-28T15:22Z) and #3 (2026-08-29T09:51Z) added .eval_results YAML metadata mirroring model-card numbers (source field: 'Model Card') — HF metadata normalization, NOT independent reproduction of the weights; trend #1 criterion (a) remains open. Downloads @2026-08-29T11:35Z: FP8 repo 8,804 (30d, effectively all-time) + 1,207 likes; BF16 repo 482. Recommendation stays TRIAL.
2026-08-31 Adoption & reproduction check @2026-08-31T00:00Z: FP8 repo downloads jumped from 8,804 (8/29 sample) to 50,116 (~5.7x in ~2 days), likes 1,207 -> 1,336 — the 753B flagship artifact is now being pulled at scale alongside the Flash variant. HF discussions #7-#9 (8/29-30) remain minor (release-cadence question, citation error, praise); still no third-party Terminal Bench / SWE reproduction of the weights — trend #1 criterion (a) unmet. Recommendation stays TRIAL.
2026-09-02 adoption & reproduction check @2026-09-02T00:0xZ: FP8 repo 50,116 -> 94,403 downloads (+88% in ~2 days, ~8 days after listing), likes 1,336 -> 1,466; BF16 repo now also public. HF discussions #10-#15 (8/31-9/1) are still metadata-sync PRs (TB 2.1/3.0 harness notes, Toolathlon-Verified eval result) and refusal complaints ('Safety Maxxed to oblivion') — no third-party Terminal Bench / SWE reproduction; trend #1 criterion (a) still unmet. Ecosystem note: Baseten's 9/1 'efficient frontier of LLM inference' blog uses GLM-5.3 as its canonical agentic-coding serving example (conceptual framework; no Flash-specific throughput numbers). Stays TRIAL
2026-09-05 Adoption & reproduction check @2026-09-05: FP8 repo 94,403 -> 303,534 downloads (+222% in ~3 days, 30d window), likes 1,466 -> 1,695; unsloth/GLM-5.3-GGUF derivative at 164,424. Discussions #16-#19 (9/2-9/4): #16 merged eval-results metadata PR (HF staff, terminal-bench-3.0 dataset pointer), #17 spam upload, #18 an independent quantization-fidelity measurement (KL vs reference on frozen tokens) — a genuine independent measurement but NOT a Terminal-Bench/SWE reproduction, #19 instruction-following complaint; trend #1 criterion (a) remains unmet. Stays TRIAL
标签
glmzhipucodingagentscybersecurityopen-weights-pending