Z.ai Open-Sources GLM-4.7, a New Generation Large Language Model Built for Real Development Workflows

SINGAPORE--(BUSINESS WIRE)--Dec 24, 2025--

Z.ai released GLM-4.7 ahead of Christmas, marking the latest iteration of its GLM large language model family. As open-source models move beyond chat-based applications and into production environments, they are increasingly expected to handle long-running tasks. GLM-4.7 has been developed with these requirements in mind.

This press release features multimedia. View the full release here: https://www.businesswire.com/news/home/20251223393714/en/

Unlike earlier systems focused on single-turn interactions, GLM-4.7 targets development environments that involve longer task cycles, frequent tool use, and higher demands for stability and consistency.

Built on GLM-4.6 with a focus on engineering use

Building on GLM-4.6, GLM-4.7 represents a clear step forward, with a design that leans more firmly toward engineering use. Support for coding workflows, complex reasoning, and agent-style execution has been strengthened, giving the model greater consistency even in long, multi-step tasks, as well as more stable behaviour when interacting with external tools. For developers, this translates into something practical: a model that can be used in everyday engineering work with greater confidence.

The improvements extend beyond technical performance. In conversational, writing, and role-playing settings, GLM-4.7 produces output that is more natural and economical, with a tone closer to everyday communication. Together, these changes point to GLM evolving into a more coherent open-source system, rather than a loose collection of related models. In an ecosystem where many projects remain fragmented or narrowly scoped, that coherence stands out.

Designed for real development workflows

As artificial-intelligence systems move beyond chat-centred applications, developers are facing a new set of challenges, and expectations for model quality have become more exacting. A capable model must do more than understand requirements or follow structured plans. It also needs to call external tools correctly and remain consistent across long, multi-step tasks. As task cycles lengthen, even minor errors can accumulate quickly, driving up debugging costs and stretching delivery timelines. GLM-4.7 was trained and evaluated with these real-world constraints in mind.

In multi-language programming and terminal-based agent environments, the model shows greater stability across extended workflows. It already supports “think-then-act” execution patterns within widely used coding frameworks such as Claude Code, Cline, Roo Code, TRAE and Kilo Code, aligning more closely with how developers approach complex tasks in practice.

Z.ai evaluated GLM-4.7 on 100 real programming tasks in a Claude Code-based development environment, covering frontend, backend and instruction-following scenarios. Compared with GLM-4.6, the new model delivers clear gains in task completion rates and behavioural consistency. This reduces the need for repeated prompt adjustments and allows developers to focus more directly on delivery. On the basis of these results, GLM-4.7 has been selected as the default model for the GLM Coding Plan.

Reliable performance across tool use and coding benchmarks

Across a range of public benchmarks related to code generation and tool use, GLM-4.7 delivers solid and competitive overall performance. On BrowseComp, a benchmark focused on web-based tasks, the model scores 67.5. On τ²-Bench, which evaluates interactive tool use, GLM-4.7 achieves a score of 87.4, the highest reported result among publicly available open-source models to date.

In major programming benchmarks including SWE-bench Verified, LiveCodeBench v6, and Terminal Bench 2.0, GLM-4.7 performs at or above the level of Claude Sonnet 4.5, while showing clear improvements over GLM-4.6 across multiple dimensions.

On Code Arena, a large-scale blind evaluation platform with more than one million participants, GLM-4.7 ranks first among open-source models and also holds the top position among models developed in China. Z.ai has emerged as a serious contender in the global open-source AI landscape, particularly in areas where reliability in real coding scenarios matters most.

More predictable and controllable reasoning

GLM-4.7 introduces more fine-grained control over how the model reasons through long-running and complex tasks. As artificial-intelligence systems move steadily into production, such capabilities have become an increasing focus for developers. The model is able to maintain consistency in its reasoning across multiple interactions, while also adjusting the depth of reasoning according to task complexity. This makes its behaviour within agent systems more predictable over time.

Whether a model can be deployed reliably at scale has become a central question for teams building production-grade AI. It is a question that Z.ai continues to examine and refine as it develops the GLM series.

Improvements in front-end generation and general capabilities

Beyond functional correctness, GLM-4.7 shows a noticeably more mature understanding of visual structure and established front-end design conventions. In tasks such as generating web pages or presentation materials, the model tends to produce layouts with more consistent spacing, clearer hierarchy, and more coherent styling, reducing the need for manual adjustment downstream.

At the same time, improvements in conversational quality and writing style have broadened the model’s range of use cases. These changes make GLM-4.7 more adaptable to creative and interactive applications, extending its role beyond purely engineering-focused scenarios.

Ecosystem integration and open access

GLM-4.7 is available via the BigModel.cn API and is fully integrated into the z.ai full-stack development environment. Developers and partners across the global ecosystem have already incorporated the GLM Coding Plan into their tools, including platforms such as TRAE, Cerebras, YouWare, Vercel, OpenRouter and CodeBuddy. Adoption across developer tools, infrastructure providers and application platforms suggests that GLM-4.7 is beginning to move beyond research settings and into wider engineering and product use.

With GLM-4.7, Z.ai continues its long-standing approach to open-source large language models: building systems that can be used reliably in real projects. The team aims to make advanced AI more practical and dependable for developers and enterprises worldwide. As open-source models take on a more prominent role in the global technology ecosystem, Z.ai’s progress offers a clear indication of how such systems may continue to evolve, and what they might enable next.

Default Model for Coding Plan: https://z.ai/subscribe

Try it now: https://chat.z.ai/

Weights: https://huggingface.co/zai-org/GLM-4.7

Technical blog: https://z.ai/blog/glm-4.7

 

GLM-4.7 ranks #6 in WebDev and is the #1 open model.

WASHINGTON (AP) — The State Department announced Tuesday it was barring five Europeans it accused of leading efforts to pressure U.S. tech firms to censor or suppress American viewpoints.

The Europeans, characterized by Secretary of State Marco Rubio as “radical” activists and “weaponized” nongovernmental organizations, fell afoul of a new visa policy announced in May to restrict the entry of foreigners deemed responsible for censorship of protected speech in the United States.

“For far too long, ideologues in Europe have led organized efforts to coerce American platforms to punish American viewpoints they oppose,” Rubio posted on X. “The Trump Administration will no longer tolerate these egregious acts of extraterritorial censorship.”

The five Europeans were identified by Sarah Rogers, the under secretary of state for public diplomacy, in a series of posts on social media. They include the leaders of organizations that address digital hate and a former European Union commissioner who clashed with tech billionaire Elon Musk over broadcasting an online interview with Donald Trump.

Rubio's statement said they advanced foreign government censorship campaigns against Americans and U.S. companies, which he said created “potentially serious adverse foreign policy consequences” for the U.S.

The action to bar them from the U.S. is part of a Trump administration campaign against foreign influence over online speech, using immigration law rather than platform regulations or sanctions.

The five Europeans named by Rogers are: Imran Ahmed, chief executive of the Centre for Countering Digital Hate; Josephine Ballon and Anna-Lena von Hodenberg, leaders of HateAid, a German organization; Clare Melford, who runs the Global Disinformation Index; and former EU Commissioner Thierry Breton, who was responsible for supervising social media rules

Rogers in her post on X called Breton, a French business executive and former finance minister, the “mastermind” behind the EU's Digital Services Act, which imposes a set of strict requirements designed to keep internet users safe online. This includes flagging harmful or illegal content like hate speech.

She referred to Breton warning Musk of a possible “amplification of harmful content” by broadcasting his livestream interview with Trump in August 2024 when he was running for president.

Breton responded Tuesday on X by noting that all 27 EU members voted for the Digital Services Act in 2022. “To our American friends: ‘Censorship isn’t where you think it is,’” he wrote.

French Foreign Minister Jean-Noël Barrot said France condemns the visa restrictions on Breton and the four others. Also posting on X, he said the DSA was adopted to ensure that “what is illegal offline is also illegal online.” He said it “has absolutely no extraterritorial reach and in no way concerns the United States.”

A statement from Ballon and von Hodenberg, the co-CEOs of HateAid, called it "an act of repression by a government that is increasingly disregarding the rule of law and trying to silence its critics by any means necessary.”

Most Europeans are covered by the Visa Waiver Program, which means they don’t necessarily need visas to come into the country. They do, however, need to complete an online application prior to arrival under a system run by the Department of Homeland Security, so it is possible that at least some of these five people have been flagged to DHS, a U.S. official said, speaking on condition of anonymity to discuss details not publicly released.

Other visa restriction policies were announced this year, along with bans targeting foreign visitors from certain African and Middle Eastern countries and the Palestinian Authority. Visitors from some countries could be required to post a financial bond when applying for a visa.

Associated Press Diplomatic Writer Matthew Lee contributed to this report.

Secretary of State Marco Rubio listens as President Donald Trump speaks at his Mar-a-Lago club, Monday, Dec. 22, 2025, in Palm Beach, Fla. (AP Photo/Alex Brandon)