GLM-5.3 · Media / benchmark · Personal experience
The hands-on report gives concrete web-generation and tool-loop observations, not a benchmark ranking.
Unverified: the original source could not be rechecked. Historical figures below are not current verified results.
APPSO received early access to GLM-5.3 for testing and evaluated it on tasks including 3D web generation, web application development, and macOS utility development. It described the model as “the strongest model in its size class and the first choice for developers who code.”
The week of its release was a “fierce clash of the gods”: Grok 4.6 (boosted by Cursor data), the official release of DeepSeek V4 Pro, Gemini Flash, and GLM-5.3 all arrived in quick succession.
The official position: coding ability is comparable to Fable 5 and GPT-5.6 Sol; internal testers reported that it felt better than Kimi K3 and DeepSeek V4-Pro-0813.
Its performance across six benchmarks was solid, with GDPVal (introduced by OpenAI) taking first place; it also ranked near the top on AutomationBench and Agents' Last Exam.
Its parameter count is comparable to that of its predecessor (roughly a 700-billion-parameter/743B base model), unlike the “trillion-parameter” path taken by Kimi K3 (2.8T) and Qwen3.8-Max. Z.ai said: “We are efficiently advancing reinforcement learning on exactly the same base as GLM-5.2. We may still be far from reaching the intelligence ceiling of this base.”
Z.ai open-sourced Slime, a post-training framework, covering the complete post-training workflow for release-grade models. It has been used since GLM 4.5 and supports GLM, some Qwen and DeepSeek models, and Llama 3.
3D interactive simulation of human blood circulation: The delivery was relatively rough (stick figures and stacked organs), but offered extensive customization, including view layers, physiological parameters such as heart rate and blood pressure, an optional hemorrhagic-shock scenario, and visible blood-flow directions. The assessment was: “It can be used in teaching, but it is not that attractive.”
Pyramid skateboard game (a DeepSeek Harness test, using GLM-5.3 + Claude Code at the highest reasoning depth): The game played well and the scene rendering was accurate. “If you do not operate it properly, you may really lose.” The downside was that the skateboard’s afterimage was too “eye-catching.”
Jellyfish lake 3D scene (millions of photorealistic jellyfish drifting in an emerald lake): “It is indeed quite beautiful.” The jellyfish were not rendered as simple circles as the human figures were; the prompt was quite long.
Claude Code compatibility issue: With automatic command approval enabled, it frequently reported “auto mode cannot determine the safety of Bash right now.” Third-party models have not yet been adapted to Claude Code’s automatic-mode mechanism. Switching to ZCode provided a better experience: like Codex, it could take screenshots, read the screen, analyze problems, and continue optimizing, while running for longer. A planet-colliding-with-Earth simulation was still repeatedly testing and checking in ZCode after more than an hour; the final effect was striking, with the crust cracking, lava erupting, and fragments forming a planetary ring.
Its performance on cybersecurity tasks was on par with Claude Mythos 5.
ExploitGym (the test OpenAI uses to attack Hugging Face when evaluating unreleased models): It completed 130 of 898 questions within the six-hour limit.
A public cybersecurity disclosure ledger (cvd.z.ai) records the model’s findings. The oldest vulnerability it found can be traced back roughly 40 years. “It had not been found for 40 years, and GLM-5.3 found it as soon as it took action.”
It is now available in Zcode (Z.ai’s official agent application) and AutoClaw (a productivity tool), with full access for GLM Coding Plan users and subscriptions open. A reset occurred once on the afternoon of the release day.
Third-party platforms including WorkBuddy, QwenWork, and TraeWork opened early access. The API is expected to open next Tuesday, and the full weights are expected to be open-sourced within two weeks.
The API pricing on the official website is still listed for GLM-5.2; at the same size, GLM-5.3 is estimated not to change much in price.
“The significance of GLM-5.3 … is that it has returned to the position that originally belonged to it: the strongest model in its size class and the first choice for developers who code.”
“K3 uses 2.8T [parameters], DeepSeek V4 uses 1.6T, while GLM 5.2 uses less than half as many parameters and achieves the same degree of intelligence.”
At a broader level, GLM-5.3 and this succession of new models are making the “shelf life of the strongest large model” shorter and shorter (GLM was overtaken by Kimi after 20 days, Kimi was overtaken by DeepSeek after 20 days, and so on).
The figures, task set, reasoning tier, and client conditions apply only to the listed source and collection snapshot. Different versions, harnesses, or providers must not be compared directly; undisclosed parameters remain unknown.
For a reproduction, fix the model version, provider or client, reasoning tier, tools, task-set version, sample count, and collection date, and record failures, retries, and human corrections. Full steps are in the source notes below.
Tencent News (republished from the official APPSO account) · APPSO · Original publication date 2026-08-14 · Site edit date 2026-09-20
Open original sourceGLM-5.3
Download the Tabbit client to check model access
