What changed in GLM-5.3
Where GLM-5.3 sits
GLM-5.3 is a large language model released by Z.ai on August 14, 2026.
The base stays put; the gains are all post-training
The unusual part is that the base model was not rebuilt. Z.ai states plainly that it uses the same base model as GLM-5.2 and that every gain comes from post-training. Post-training is the later stage in which a finished model is made to solve tasks and adjusted repeatedly.
"It uses the same base model as GLM-5.2 — every gain comes from post-training."(opening)／"Over the past month we kept scaling on this stack: more environments, more diverse tasks, and more compute spent training on them."(preceding paragraph) — from the Z.ai official blog
Coding improved 50% on the in-house benchmark
Z.ai states that coding improved 50% over 5.2 on its in-house Z.ai Code Bench, and claims it is the most capable open-weights model for coding. Open weight means the model's weights — the numbers produced by training — are distributed so you can run it yourself. For the previous generation, see the GLM-5.2 explainer.
| Benchmark | GLM-5.3 | GLM-5.2 | Mythos 5 (listed as Fable 5 w/ fallback) | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal Bench 3.0 | 28.3 | 4.6 | 33.7 | 34.6 |
| DeepSWE v1.1 | 66.9 | 46.2 | 69.7 | 72.7 |
| CyberGym | 84.5 | 77.2 | 83.8 | 83.6 |
| ExploitBench | 54.4 | 24.4 | 78.0 | 76.5 |
| Agents' Last Exam | 28.5 | 23.8 | 23.8 | 28.6 |
"Stronger Coding: GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench."(summary bullets)／"GLM-5.3 improves from 4.6 to 28.3 on Terminal-Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1, and from 23.8 to 28.5 on Agents' Last Exam."(Stronger Coding section) — from the Z.ai official blog
A jump in cyber capability, and the gap that remains
CyberGym (vulnerability discovery, %)
The top three sit within 0.9 points, so the lead is marginal. The 7.3-point gain over its own previous generation is larger than the gap to any competitor.
The area the company describes as having grown faster than expected is cyber. Closed here means commercial models that do not publish their weights.
84.5% on CyberGym
On CyberGym, which measures vulnerability discovery, GLM-5.3 scored 84.5%. Z.ai calls this the best result on the benchmark, ahead of Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%. CyberGym starts from white-box source code and tests whether a model can identify vulnerabilities and validate them by triggering faults.
"On CyberGym, which starts from white-box source code and tests whether the model can identify and validate vulnerabilities by triggering faults, GLM-5.3 scores 84.5%, up from GLM-5.2's 77.2% — the best result on the benchmark, ahead of Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%)."(Emergent Cyber Capability section) — from the Z.ai official blog
Further up the exploitation chain, the gap persists
On ExploitBench, which covers deeper stages of an attack, it reaches 54.4%. That more than doubles 5.2's 24.4%, but it does not reach Mythos 5's 78.0% or GPT-5.6 Sol's 76.5%. Z.ai summarizes the pattern: the further up the exploitation chain a benchmark sits, the larger the gain — and the wider the remaining gap to closed models.
"On ExploitBench, which requires deeper reasoning about real vulnerabilities and their exploitation, GLM-5.3 reaches 54.4%, more than doubling GLM-5.2's 24.4%, while Mythos 5 and GPT-5.6 Sol score 78.0% and 76.5%, respectively."／"The pattern across the three is consistent: the further up the exploitation chain a benchmark sits, the larger the gain from GLM-5.2 — and also the wider the remaining gap to the closed frontier."(both from the Emergent Cyber Capability section) — from the Z.ai official blog
Where to use it, and when the weights arrive
Availability
The only thing being held back is the weights.
Entry points today, and the weight release schedule
GLM-5.3 is already available through the API and has been rolled out to all GLM Coding Plan users. You can try it from your preferred coding agent, including ZCode, Claude Code, and OpenCode. ZCode is Z.ai's own coding environment. What is staged is the weights, with a stated target of two weeks after launch. In the API, disabling thinking is no longer supported and reasoning_effort now takes three levels — low, high, and max — so existing implementations need migration.
"Try GLM-5.3 in your favorite coding agents—ZCode, Claude Code, OpenCode, and more."／"We've rolled out GLM-5.3 to all GLM Coding Plan users."(both from the Use GLM-5.3 with GLM Coding Plan & ZCode section)／"GLM-5.3 supports three thinking effort levels: low, high, and max. Disabling thinking is no longer supported by GLM-5.3."／"Migration required: If your application currently uses thinking.type: \"disabled\", change it to enabled and set reasoning_effort to low before updating the model ID to glm-5.3. Otherwise, the request will fail."(both from the API Changes in GLM-5.3 section)／"Open Source: We will release the weights in two weeks after launch, once safety evaluation and hardening are complete."(summary bullets) — from the Z.ai official blog
Findings against real code
Working with security teams in China and running it against real code, the company states that after expert verification, triage, and deduplication, the model identified 2,436 vulnerabilities across 269 projects, of which 1,097 were medium-to-high severity. Some had gone unnoticed for years or decades, with the oldest dating back roughly 40 years.
"the model identified 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues."／"Many had remained unnoticed for years or even decades, with the oldest dating back roughly 40 years."(both from the Emergent Cyber Capability section) — from the Z.ai official blog
To convert an English technical blog into markdown with its structure intact, the following tool can help.
