Z.ai · 2026-08-14 · major
GLM-5.3 — Z.ai's coding model improves without retraining the base
GLM-5.3 keeps the same base model as GLM-5.2 and gets all of its gains from more post-training. Z.ai reports Terminal-Bench 3.0 rising from 4.6 to 28.3 and says cyber skill grew faster than expected.
Z.ai got a large jump in coding and security skill out of GLM-5.2's base model by training it harder after the fact.
Key specs
| Base | Same as GLM-5.2 |
|---|---|
| Weights | ~2 weeks after launch |
Quick facts
| Maker | Z.ai (Zhipu AI) |
|---|---|
| Base model | Same base as GLM-5.2, not retrained |
| What's new | Scaled post-training only |
| Thinking effort | Low, high or max — cannot be turned off |
| Availability | Z.ai API, GLM Coding Plan, ZCode |
| Open weights | Promised about two weeks after launch |
| Security findings | 2,436 vulnerabilities, 53 CVEs public |
Benchmarks
What is it?
GLM-5.3 reuses the GLM-5.2 base model and takes every capability gain from scaled-up post-training instead of a new pretraining run. Z.ai calls it the strongest open-weights coding system it has measured, and says the model's cybersecurity skill grew faster than the company expected as training scaled. Thinking is now always on, with three effort levels to pick from.
How does it work?
Rather than pretraining a new base, Z.ai scaled the post-training stage on top of the GLM-5.2 base and pushed hard on long-horizon agent work — the kind of task where a model runs a terminal for many steps instead of writing one file. That is where the gap shows up: the longest-horizon coding benchmarks move the most, while short single-shot tasks move far less. Callers choose low, high or max thinking effort, but cannot switch thinking off.
Why does it matter?
A post-training-only upgrade is much cheaper than a new base model, so GLM-5.3 suggests labs can keep buying real capability without another pretraining bill. The security side cuts both ways: Z.ai credits the model's vulnerability hunting with 2,436 findings across 269 open-source projects, which is useful defensive work and also a sign that offensive skill is arriving in models anyone can rent by the token.
Who is it for?
developers on coding plans, agent builders, security researchers
Frequently asked questions
- Are the GLM-5.3 weights available to download?
- Not at launch. Z.ai says it will publish the GLM-5.3 weights roughly two weeks after release, once safety evaluation and hardening finish. Until then the model is API-only through Z.ai. The existing GLM-5 repository on GitHub, where earlier weights in this series landed, carries an Apache-2.0 license.
- How does GLM-5.3 compare to Claude and GPT models?
- Unite.AI reports that GLM-5.3 edges past Claude Opus 4.8 when thinking effort is matched, but still trails Claude Fable 5 and GPT-5.6 Sol on the harder coding benchmarks. Z.ai's own framing is narrower: it claims GLM-5.3 is the strongest open-weights coding system the company has measured, not the strongest model overall.
- Where can I use GLM-5.3 right now?
- GLM-5.3 is live on the Z.ai API, inside the GLM Coding Plan, and in ZCode, Z.ai's official coding harness. Existing GLM Coding Plan subscribers were moved onto it automatically, so no separate purchase is needed. Z.ai has not published a per-token price row for GLM-5.3 yet.
- What did Z.ai's security work with the model actually find?
- Z.ai credits its GLM-5.3 security effort with 2,436 vulnerabilities across 269 open-source projects, 1,097 of them rated critical or high. Its public disclosure ledger at cvd.z.ai lists 53 of those as publicly disclosed CVEs, with 2,383 still under embargo, covering projects such as Linux, Safari, FreeBSD and GStreamer.
- Can I turn off thinking in GLM-5.3?
- No. GLM-5.3 removes the option to disable thinking that earlier GLM releases offered. Instead, callers pick one of three thinking effort levels — low, high or max — and trade latency and token spend against quality. Teams that relied on a zero-thinking fast path will need to move to the low effort level.
Try it
Call model id glm-5.3 on the Z.ai API, or use it inside the GLM Coding Plan and ZCode.