It's always hit & miss when it's post-training based. I.e. gpt 5.1/5.5 and opus 4.8 all had regressions on functionality out of the training scope (or well, for 4.8 I guess the training wasn't great to begin with.) I'll check it out when it's open.

Funnily, for 5.2 my best harness is plain codex-oss with gpt prompts and they don't even mention that here. So I'll probably have to do a bunch of tests again.

5.2 "felt" better than the benches,  interested to give this one a play

Grok 4.6 was feeling a lot better than 4.5 for a few days, fast and not lazy, but yesterday started getting slow and lazy again... I swear Cursor is A/B testing on us constantly.

justin_shocknet

zuspotirko

Scaling post-training is all we did for GLM-5.3.

 for efficient long-context processing, 

 for large-scale asynchronous training — all running on the long-horizon task environments we have been accumulating. Over the past month we kept scaling on this stack: more environments, more diverse tasks, and more compute spent training on them.

Today we are releasing GLM-5.3. It uses the same base model as GLM-5.2 — every gain comes from post-training. Compared with GLM-5.2, it is much better at complex coding and long-horizon tasks:

 GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench. It also achieve open-source SOTA on public benchmarks including Terminal Bench 3.0 and Agents' Last Exam.

 As we scaled post-training, cyber capability developed faster than we expected. GLM-5.3 is state of the art on CyberGym for vulnerability discovery, and its gains are largest further up the exploitation chain, where it more than doubles GLM-5.2 on exploitation benchmarks.

 We will release the weights in two weeks after launch, once safety evaluation and hardening are complete.

> **Scaling post-training is all we did for GLM-5.3.** With GLM-5.2 we built the stack: [IndexShare](https://arxiv.org/abs/2603.12201) for efficient long-context processing, [SAO](https://arxiv.org/abs/2607.07508) for RL on long-horizon tasks, and [slime](https://github.com/THUDM/slime) for large-scale asynchronous training — all running on the long-horizon task environments we have been accumulating. Over the past month we kept scaling on this stack: more environments, more diverse tasks, and more compute spent training on them.
>
> Today we are releasing GLM-5.3. It uses the same base model as GLM-5.2 — every gain comes from post-training. Compared with GLM-5.2, it is much better at complex coding and long-horizon tasks:
>
> **Stronger Coding:** GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench. It also achieve open-source SOTA on public benchmarks including Terminal Bench 3.0 and Agents' Last Exam.
>
> **Emergent Cyber Capability:** As we scaled post-training, cyber capability developed faster than we expected. GLM-5.3 is state of the art on CyberGym for vulnerability discovery, and its gains are largest further up the exploitation chain, where it more than doubles GLM-5.2 on exploitation benchmarks.
>
> **Open Source:** We will release the weights in two weeks after launch, once safety evaluation and hardening are complete.
>
>
> ![](https://m.stacker.news/152204)

***

Notable:

- No new tech, just more training
- Staggered release for "safety"
- Beats commercial models on a bunch of their own self-eval benches.