Zishan Lee of Z.ai dialed in remotely to introduce GLM 5.2 (and GLM 4.2), the company's latest open-weight models. He explained that GLM stands for General Language Model — a name rooted in a 2021 paper on autoregressive blank-filling — and positioned GLM 5.2's coding and agentic capability as sitting between Opus 4.7 and 4.8, with a new "high" thinking level for better token efficiency. Lee argued the case for open weights (security, on-prem control, fine-tuning, co-designing the future) and closed with a "one more thing": Zcode, Z.ai's own Codex-like coding harness.
Lee opened by clearing up confusion between Z.ai and GLM: the model name is not a brand invention but a generic term, General Language Model. It traces to a 2021 paper on training with autoregressive blank filling, making Z.ai one of the first labs exploring large models alongside OpenAI, Anthropic, and DeepMind.
Although the company no longer uses the original GLM architecture, it retains the GLM name as a brand across releases like 5.1 and 5.2, which have become its broadest product and model line.
GLM 5.2 (referred to alongside GLM 4.2) specializes in coding and agentic tasks. Addressing rumors about where it lands versus frontier models, Lee showed benchmarks placing it between Opus 4.7 and 4.8, using the hardest problems such as "deep tunnel bench 2.1" and long-horizon tasks, where it is on par with at least Opus 4.7 and a significant improvement over the prior generation.
For the first time the team added a "high" thinking level, motivated by harder tasks consuming more tokens and a focus on token efficiency. Notably, the non-thinking model already outperforms the GLM 5.1 thinking model — which Lee framed as a huge leap for open-weight models.
Lee stressed GLM is more than a coding model. Beyond its use inside Claude Code, Codex, and open-code, Z.ai trained heavily on GDPval, math, role play, and general chat, aiming to improve every aspect of the model.
He pointed to the Artificial Analysis Intelligence Index, where GLM leads other open-weight models by a wide margin and sits close to frontier models, and encouraged the audience to use it for daily workflows and general chat, not just coding.
Addressing why the company open-weights its models rather than protecting business against inference providers, Lee cited both user and company needs. Enterprises and governments — especially in the western world — want security, control, and trust, so Z.ai uploads to Hugging Face for on-premise use.
Open weights also enable diversity through fine-tuning in domains like legal, finance, and security. Lee noted Harvey is fine-tuning GLM 4.1 and may move to 5.2, and that other companies see fine-tuning GLM as a differentiation strategy. Sharing architecture and training recipe lets customers co-design and co-shape the future with the open-source community.
Lee directed the audience to Z.ai's tech blog, which covers the Hugging Face repo, the chatbot agent, the API, a coding-plan subscription (like Codex or Claude Code), and the training pipeline and recipe explaining why the model performs well.
As a closing surprise, he unveiled Zcode — Z.ai's own coding harness, built for GLM 4.2 but supporting all frontier models with bring-your-own-key, operating similarly to Codex. The host added that Z.ai holds four of the top eight Hugging Face contributor slots present at the fair, and that the ecosystem includes Nvidia, Unsloth, and Ollama.
"GLM actually represent general language model for training with auto regressive blank fill and that paper was published back in 2021"
"we no longer use GLM as the architecture we still use the name GLM as our brand name"
"you can see it's somewhere between Opus 4.7 and 4.8"
"even without thinking the non-thinking model is better than the 5.1 thinking model"
"GLM is more than coding model because people use it inside Claude Code, Codex, OpenCode but actually we have trained a lot of things outside coding"
"it's the first time we share Zcode to the whole community"
| Time | Topic |
|---|---|
| 02:04 | What GLM means: General Language Model, 2021 paper |
| 03:05 | Intelligence upper bound beyond IQ |
| 04:36 | Adding a 'high' thinking level and token efficiency |
| 05:38 | More than coding: GDPval, math, role play |
| 06:39 | The case for open weights |
| 10:44 | One more thing: Zcode harness |