GLM-5.3
Zhipu AI·
Overview
Post-training-only update to GLM-5.2: Z.ai leaves the 753B MoE base model unchanged and attributes the capability gains entirely to scaled post-training. Weights landed on Hugging Face on 2026-08-28, fourteen days after launch, once Z.ai's safety review completed — but under a bespoke GLM-5.3 License instead of the MIT license GLM-5.2 shipped under: use, modification, distribution, sublicensing, sale, deployment and fine-tuning are all permitted, but a company with more than $10B aggregate revenue over any 12 consecutive months must pass a Z.ai security review before hosting the model commercially. Individual users and smaller companies are unaffected. Z.ai reports its results on a new generation of agentic benchmarks rather than the axes tracked here (maximum thinking effort): Terminal-Bench 3.0 28.3 (GLM-5.2: 4.6), DeepSWE v1.1 66.9 (46.2), Agents' Last Exam 28.5 (23.8), CyberGym 84.5, plus AutomationBench, GDPVal-AA v2 and HLE with tools. Only Terminal-Bench 3.0 has a column here (its own axis, separate from the 2.x scale); the rest do not map onto the tracked benchmarks, and Z.ai published no GPQA, MMLU-Pro or SWE-bench Pro table.
Capabilities and innovations
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| Terminal-Bench 2.x | 71.54% | independent Terminal-Bench 2.1 71.54%, measured by Vals AI with the Terminus 2 harness (accessed 2026-09-23). Vals AI is the reference harness for this field so values compare across vendors; vendor and other third-party figures stay in the notes below. No Terminal-Bench 2.x value was recorded before. |
| Terminal-Bench 3.0 | 32.4% | independent Terminal-Bench 3.0 32.4% (± 2.9, 95% CI), official leaderboard maintained by the Terminal-Bench team (Stanford / Harbor / Laude Institute), max effort, agent harness: Claude Code (accessed 2026-09-23). Z.ai self-reported 28.3 in the GLM-5.3 launch table; the leaderboard value is kept as primary per the vendor-vs-independent rule. Do not compare against the terminal_bench column: 3.0 is a harder generation on a different scale. |
| Terminal-Bench 4.0 | 41.8% | independent Terminal-Bench 4.0 41.8% (± 3.2, 95% CI), official leaderboard maintained by the Terminal-Bench team (Stanford / Harbor / Laude Institute), max effort, agent harness: Claude Code (accessed 2026-09-23). 4.0 is its own generation and not comparable with the 2.x or 3.0 fields. |
| DeepSWE v1.1 | 66.9% | vendor DeepSWE v1.1 66.9 from Z.ai's GLM-5.3 launch table at maximum thinking effort, reported in place of SWE-bench Pro (GLM-5.2: 46.2 in the same table). Vendor self-report over its own model. Datacurve's own DeepSWE board was not reachable from this environment, so the value is not cross-checked against it. |
- GPQA Diamond: independent evaluation pending — Z.ai published no GPQA table for this release; Vals AI and Artificial Analysis had no GLM-5.3 entry as of 2026-08-17.
- Humanity’s Last Exam (no tools): not reported by the vendor — Z.ai reports HLE with tools only, which is not comparable to the no-tools HLE used here (same handling as the GLM-5.2 entry).
- SWE-bench Pro: not reported by the vendor — Z.ai reports DeepSWE v1.1 (66.9, GLM-5.2: 46.2, both now recorded in deepswe_v1_1) instead of SWE-bench Pro.
- MMLU-Pro: independent evaluation pending — Not in Z.ai's table; Vals AI had no GLM-5.3 entry as of 2026-08-17.
- MMMU-Pro: not reported by the vendor
Architecture and hardware
- Parameters
- 753B total, 40B active per token (MoE)
- Estimated VRAM at Q4
- ~433 GB, Frontier class
Reliability
- tau3_banking
- 50.3 ()third party
- omniscience_index
- 14.3 ()independent
Links
More from Zhipu AI
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.