Z.ai Ships GLM-5.3 as Open-Weight, Matching Closed Flagships on Coding

Z.ai publishes GLM-5.3 as open-weight

On August 28, Z.ai released GLM-5.3 under an open-weight license on Hugging Face, and the model quickly climbed to the top of Hacker News. The release is notable for a simple reason: GLM-5.3 uses the same base model as GLM-5.2. According to the model card, every gain comes from post-training.

That matters because it shows how much headroom remains after pretraining. Instead of a bigger model, Z.ai shipped a better-trained one.

Coding and long-horizon agents

The company calls GLM-5.3 "the most capable open-weights model for coding," with a 50 percent improvement over GLM-5.2 on its in-house Z.ai Code Bench. It also claims open-source state of the art on Terminal Bench 3.0 (28.3 versus 4.6) and Agents' Last Exam, benchmarks built for long, multi-step tasks rather than single prompts.

On public leaderboards the model now sits in the same band as closed flagships such as Opus 4.8 and GPT-5.6 Sol, alongside other Chinese open releases like Kimi K3 and DeepSeek-V4 Pro.

An unexpected cyber capability

The most discussed detail is security. Z.ai reports that cyber capability "developed faster than we expected" as post-training scaled: GLM-5.3 tops CyberGym for vulnerability discovery and more than doubles GLM-5.2 on exploitation benchmarks (105 versus 29 at two hours). That is a dual-use headline — the same model that finds bugs can help write exploits, and anyone can download and run it locally.

For teams, the practical takeaway is twofold: frontier-adjacent coding agents are increasingly free to self-host, with support already in SGLang, vLLM, Transformers and Unsloth plus a reasoning_effort switch for cost control. At the same time, security teams should assume capable offensive tooling is now one download away.

📖 Source