GLM-5.3
Released Aug 14, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Type
- LLM
- Modality
- Text
- Context window
- 1M
- Max output
- 128K
- Speed (max effort)
- 72 tok/s · 1.63s to first answer token
- Price ($/MTok in / out)
- $1.40 / $4.40
- Cost per task
- —
- Open weights
- Yes
Benchmarks2 of 10 reported
Terminal-Bench 2.1
LMArena Elo
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
AIME (latest)retired
Humanity's Last Exam
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Independently drives front-end builds, bug fixes and refactors through repeated self-verification to a deliverable state with little hand-holding
- Finds and triages real vulnerabilities at scale — Z.ai's own evaluation run surfaced thousands of flaws across public repositories
- Reaches a meaningfully higher agentic ceiling than 5.2 without a larger base model — the entire gain came from more post-training, not more parameters
Weaknesses
- Still a step behind the closed frontier on the single hardest coding problems, even as it closes most of the gap on everyday agent work
- Good at finding exploitable bugs but weaker at writing the exploits itself — defensive security strength doesn't carry over to offensive tasks
- Reasoning is always-on with only a low/high/max budget dial, and Artificial Analysis measures it at max — every call carries thinking overhead you cannot switch off
Same ~744B-parameter (~40B active) MoE base as GLM-5.2 — Z.ai's own framing is 'scaling post-training is all we did for GLM-5.3' — so predecessorId is left null per this site's rule against lineage/build-upon language as replacement evidence, even though several outlets call GLM-5.2 its 'predecessor.' Z.ai's launch benchmark table is entirely agentic/coding/cyber (Terminal-Bench 2.1 88.2%, Terminal-Bench 3.0 4.6%→28.3%, DeepSWE v1.1 46.2%→66.9%, SWE-Marathon v1.1 19.4%→42.5%, AutomationBench 26.2%→48.2%, Agents' Last Exam 23.8%→28.5%, CyberGym 84.5% — ahead of Claude Mythos 5's 83.8% and GPT-5.6 Sol's 83.6% by Z.ai's account) and doesn't touch the site's tracked knowledge/reasoning keys, consistent with the unchanged-base-model claim. Context window and max output are carried over unchanged from GLM-5.2 (1M / 128K) pending a dedicated 5.3 model card. Weights have since been published on Hugging Face (zai-org/GLM-5.3) under a bespoke "GLM-5.3 License" rather than the MIT terms used for GLM-5.2 and GLM-5.3-Flash: an MIT-style grant covering weights, code and derivative works, with one added condition — any Model-as-a-Service operator whose group revenue exceeds US$10B in a 12-month period must pass a Z.ai security review before commercial use. Z.ai's API now lists the model as glm-5.3 at $1.40/$4.40 per M input/output tokens (cached input $0.26); Artificial Analysis measures the max-effort variant at 72 tok/s and 1.63 s to first answer token, but its per-task cost figure could not be read. The exact weights-publication date (reported as 2026-08-28 by several trackers) could not be confirmed from a primary page. LMArena Elo 1483 is the "glm-5.3-max" text-arena listing (arena.ai, 2026-09-13 update, rank 19).
For developers
API model strings
- zhipu
glm-5.3
Licence
GLM-5.3 License · Restricted · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor