Watch on YouTubeGLM-5.2 reaches the level of at least Claude Opus 4.7 on hard, long-horizon coding and agentic tasks, according to Zixuan Li.
Z.ai releases open weights so enterprises and governments can run GLM on their own infrastructure and teams can fine-tune it for fields such as law, finance, and security.
Z Code is Z.ai's coding harness for GLM-5.2, with support for other frontier models and bring-your-own API keys.
Zixuan Li introduces GLM-5.2 and explains why Z.ai treats GLM as more than a coding model. He says the model performs at least as well as Claude Opus 4.7 on difficult long-horizon coding and agentic benchmarks, while its non-thinking mode exceeds GLM-5.1 with thinking enabled. GLM-5.2 also targets reasoning, maths, general chat, role play, and other uses. Li's main argument for open weights is practical. Enterprises and governments can run the model on their own servers, while companies can fine-tune it for areas such as law, finance, and security. Customers and open-source developers can also inspect the model and help shape future versions. He points listeners to Z.ai's technical blog, API, chatbot agent, coding plan, and training details. He closes by introducing Z Code, a coding harness built for GLM-5.2 that also connects to other frontier models.
Zixuan Li explains that the company is called Zhipu and that GLM originally referred to General Language Model pre-training with auto-regressive blank filling. The paper was published in 2021. Z.ai no longer uses that architecture, but it kept GLM as the model brand, including GLM-5.1 and GLM-5.2. Li places the work among early large-language-model research alongside OpenAI, Anthropic, and DeepMind.
Li says intelligence should not be reduced to IQ, AIME, or physics questions. Across GLM-4.5 to GLM-4.7, Z.ai explored reasoning, coding, and agentic abilities. This frames GLM-5.2 around work that requires a model to keep solving a task over a longer period, rather than only producing a correct answer to an isolated benchmark question.
Li places GLM-5.2 between Claude Opus 4.7 and 4.8 in the comparisons he presents. He says the evaluation uses difficult problems, including DeepSweep and Terminal-Bench, along with long-horizon tasks. Across those benchmarks, he describes the model's capabilities as at least on par with Opus 4.7 and as a significant improvement over GLM-5.1.
Z.ai added a high thinking level because harder tasks can require more tokens. Li says the team also cares about token efficiency. He makes a stronger comparison between versions: GLM-5.2 without thinking performs better than GLM-5.1 with thinking enabled. He presents this as a major improvement for an open-weight model.
Li says people use GLM inside coding tools, but coding is only one part of the model's training. Z.ai also worked on GDPval, maths, role play, and general chat. He points to the Artificial Analysis Intelligence Index, where he says GLM leads other open-weight models and is close to frontier models. His invitation is to use GLM for everyday workflows and general conversation as well as code.
Li gives security, control, and trust as reasons to release the weights. Enterprises and governments, especially in Western markets, may want to run a model on their own premises. Publishing the model on Hugging Face makes that possible. For Li, this is also a way for the wider ecosystem to examine and use a near-frontier model outside a hosted service.
Open weights allow teams to adapt GLM for domains such as legal work, finance, and security. Li names Harvey as a company fine-tuning GLM-4.1 and considering GLM-5.2. He says other companies are also considering GLM as a way to differentiate their applications. The value comes from changing the model for a particular domain instead of using the same general model everywhere.
Li says customers and individuals may need to inspect the model architecture and the recipe used to train it if they want to help shape future systems. Z.ai wants to co-design that future with customers and the open-source community. He credits model users, software projects, individual developers, and application builders with helping improve GLM.
Li introduces Z Code as Z.ai's own harness, released to the wider community for the first time. It was built for GLM-5.2, but it also supports other frontier models. Users can bring their own API key and try coding techniques such as Go or compact techniques. Z.ai is showing the harness at its booth and making it available through its online channels.
"Actually, we open the weights for several things because there are users' needs and there are our own needs."06:57