Skip to main content

Zhipu's GLM-5.3: Same Base, Bigger Brains—Open-Source Coding Crown Shifts

Zhipu AI dropped a surprise today with the release of GLM-5.3. Unlike the usual upgrade cycle, the base model hasn't changed a bit. All the magic happened in the post-training phase. By scaling up long-range task environments tenfold, diversifying environment types, and extending training time, Zhipu managed to push the same base to new heights. Internal tests show the new model's coding skills are about 50% better than GLM-5.2, making it the strongest open-source contender for programming right now.

Public benchmarks tell a similar story. Terminal-Bench 3.0 scores jumped from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Agents' Last Exam from 23.8 to 28.5. GDPval-AA v2 hit 1769 points. These numbers put GLM-5.3 in the same league as Claude Fable 5, and it's now leading the open-source pack in both Terminal Bench 3.0 and Agents' Last Exam (CLI).

Image

Post-Training Scaling: The Untapped Potential of Base Models

Zhipu's team is quick to point out that all these gains come from post-training, not from tweaking the base. Using their IndexShare, SAO, and the evolving Slime framework, they ran reinforcement learning on the same base as GLM-5.2. They even admit, "we might be far from fully exploring the intelligence ceiling of this base." That's a bold statement, and it sends a clear message: in the race for better AI, you don't always need bigger models. Sometimes, you just need to squeeze every drop of potential out of what you already have.

Image

Security First, Weights Later

On the security front, GLM-5.3 holds its own against Mythos 5 in tasks like white-box code review and vulnerability detection. That makes it a promising tool for cyber defense. But Zhipu isn't rushing to open the floodgates. They'll release the full model weights two weeks after launch, but only after thorough security assessments and hardening. The goal is to limit any potential for misuse while keeping the defensive value intact.

The Shift from Talk to Action

GLM-5.3 is already available on Zhipu's own tools—ZCode, AutoClaw, and the GLM Coding Plan—and early access is rolling out on platforms like Trae, Kouzi, WorkBuddy/CodeBuddy, and Qoder. As open-source models inch closer to their closed-source rivals in coding, the competition among domestic AI giants is changing. It's no longer about who sounds smarter in a demo. It's about who actually delivers better results in the real world.

Key Points

  • Base model unchanged: All improvements come from post-training scaling.
  • Coding boost: 50% better than GLM-5.2, now the top open-source model for programming.
  • Benchmark leaps: Terminal-Bench 3.0 up from 4.6 to 28.3; DeepSWE v1.1 up from 46.2 to 66.9.
  • Security measures: Weights released two weeks after launch, post-assessment.
  • Availability: Now on ZCode, AutoClaw, GLM Coding Plan, and early access on other platforms.