China's Z.AI Ships GLM-5.3, Calling It the High Open-Weight Coding Mannequin – Decrypt

Chinese language AI lab Z.ai launched GLM-5.3 on Thursday, a large coding mannequin it is pitching because the strongest open-weights coder in the marketplace. The mannequin is reside now via the GLM Coding Plan subscription and ZCode, with API entry and downloadable weights following after a security evaluate.

“Scaling post-training is all we did for GLM-5.3,” the corporate wrote in its launch put up. “With GLM-5.2 we constructed the stack… Over the previous month we stored scaling on this stack: extra environments, extra various duties, and extra compute spent coaching on them.”

The staff targeted extra on token effectivity, not uncooked dominance. GLM-5.3 stands at 743 billion parameters and consumes rather a lot much less tokens per job than its predecessor. Parameters are the quantity of dials a mannequin deal with whereas processing info whereas tokens are the essential unit of data a mannequin can eat or generate.

Z.ai says GLM-5.3 clears 34.5% on its in-house Z.ai Code Bench at Max effort whereas burning roughly 75,000 output tokens per job, towards GLM-5.2’s 23.4% at 96,000. Towards closed fashions, the weblog notes it beats Claude Opus 4.8 on token financial system however “stays behind Claude Fable 5, which reaches 39.5% at Max effort.”

By way of coding, GLM5.3 is an excellent performer, beating fellow Chinese language mannequin Kimi K3 on essentially the most related benchmarks.

On Terminal Bench 3.0—a take a look at of autonomous shell/instrument use in actual Linux environments—GLM-5.3 scores 28.3, barely behind closed fashions Fable 5 (33.7) and GPT-5.6 Sol (34.6). On DeepSWE v1.1, a benchmark for fixing actual GitHub points end-to-end, open rival Kimi K3 (67.5) and Fable 5 (69.7) each beat GLM-5.3’s 66.9.

The sample could be roughly summed up like this: GLM-5.3 clears its personal predecessor and a few open friends, however closed U.S. fashions nonetheless lead the headline coding boards.

The cybersecurity outcomes present one other vital leap. GLM-5.3 leads CyberGym at 84.5% and greater than doubles GLM-5.2 on exploitation benchmarks. Z.ai says the mannequin flagged 2,436 vulnerabilities throughout 269 open-source initiatives, 1,097 of them medium-to-high severity.

“GLM-5.3 takes agentic coding to the following degree, delivering a dramatic enchancment over GLM-5.2 whereas reaching higher outcomes with fewer output tokens,” Z.ai posted on X. “GLM-5.3 is obtainable now via GLM Coding Plan and ZCode. API entry and open weights will probably be launched in levels following rigorous security evaluations.”

On value, the hole with U.S. frontier fashions is the open-weights draw. Z.ai’s GLM Coding Plan runs on a factors quota (off-peak calls price half), with Zhipu’s API priced at roughly a tenth of U.S. frontier per-token charges—GLM-5.2’s official price was $1.40 in / $4.40 out per million tokens. That stacks towards GPT-5.3-Codex at $1.75 / $14 and Claude Opus 4.8 close to the highest of Anthropic’s tiers.

Z.ai is a Beijing lab included on the U.S. Entity Checklist, which suggests American companies can not export managed tech to it. Regardless of this, GLM is an especially common mannequin and Chinese language open-weight fashions already beat American ones on OpenRouter token utilization.

GLM-5.3 weights are set for public launch in about two weeks, per the launch put up—the open-weights label applies to what’s coming, not what’s downloadable in the present day.