GLM-5.3: Zhipu's open-weights model leads on agentic coding
Zhipu AI (brand Z.ai) unveiled GLM-5.3 on August 14, 2026 — the new flagship of its open GLM line. The open weights followed a risk review around August 28, 2026 on Hugging Face. The point of this release isn’t another freely available model, but where it lands: on agentic coding and security benchmarks, GLM-5.3 reaches the top among open models and closes in on proprietary frontier models.
What's new
- Release: August 14, 2026; open weights on Hugging Face after a risk review around August 28, 2026.
- Architecture: Mixture-of-Experts with roughly 744B parameters, about 40B active per forward pass. Context window: 200,000 tokens.
- Coding: around 50 percent better than GLM-5.2 on its own benchmark; Terminal-Bench 3.0 from 4.6 to 28.3 — rank one among open models.
- Agents/security: Agents’ Last Exam (CLI) 28.5 (best open-weight score, just behind GPT-5.6 Sol at 28.6); CyberGym 84.5 percent, ExploitBench 54.4 percent.
What happened
Zhipu describes GLM-5.3 with the tagline “Scaling post-training is all we did for GLM-5.3.” So the gain over GLM-5.2 doesn’t come from a new architecture but from heavier post-training. The result is a model with a clear focus: coding and cybersecurity, measured on agentic benchmarks — tasks where the model calls tools, evaluates intermediate results and keeps working across multiple steps.
The most visible jump is Terminal-Bench 3.0, which measures working with a command line: from 4.6 to 28.3. On Agents’ Last Exam (CLI), GLM-5.3 sits at 28.5, just behind GPT-5.6 Sol (28.6) — the best score among open models. The security benchmarks CyberGym (84.5 percent) and ExploitBench (54.4 percent) show the same emphasis.
Why it matters
Until now, the domain of long, agentic coding runs sat firmly with the expensive, closed frontier models. An open model that reaches GPT-5.6 Sol on exactly these tasks shifts the maths for anyone who wants to — or has to — self-host, whether for data protection, cost or sovereignty reasons.
The price for that is infrastructure: running the full weights reportedly needs around eight GPUs. GLM-5.3 isn’t a laptop option — it’s for operators with their own hardware or a rented GPU cluster. Those who can carry that effort get one of the most capable open alternatives to proprietary top models, with full control over weights and data.
One caveat remains: the benchmark figures are largely Zhipu’s own measurement. Independent reproductions for Chinese open-weight releases typically follow a few weeks later and tend to read more soberly than the launch announcement.
What you can do now
If you self-host coding agents or terminal workflows: test GLM-5.3 against your current open model on a real task, not on the benchmark list — and put the honest GPU cost into the calculation.
If data sovereignty is your criterion: GLM-5.3 belongs on the shortlist the moment self-hostable models are on the table for agentic coding. The details on architecture, licence and benchmarks are in the glossary entry → GLM-5.3, the read on the line as a whole in the hub → GLM.