People queue to have their laptops install with OpenClaw, an open-source AI assistant at the Baidu headquarter in Beijing on March 11, 2026. (Photo by ADEK BERRY / AFP via Getty Images)
AFP via Getty Images
A very peculiar scene has unfolded recently in China. Long lines of everyday citizens are lining up to have an open-source AI tool installed on their laptops. OpenClaw, a framework for building autonomous and persistent AI agents, has exploded into a national phenomenon.
Deemed “raising a lobster” which comes from the red lobster logo used by OpenClaw and users joking that installing and training the agent is like “raising” a digital pet lobster, it is transforming China’s AI landscape faster than anyone could have foreseen.
Zhipu AI (known internationally as Z.ai) is at the center of this frenzy. Zhipu, a Tsinghua University spinoff, is China’s first publicly listed 100% AI company. On Feb. 11, it launched GLM-5, which is a 744-billion parameter mixture-of-experts model. As shown in the figure below, the GLM-5 model performs competitively with western frontier models such as Anthropic’s Claude on agentic benchmarks including SWE-bench (77.8%), GPQA-Diamond (86.0%) and BrowseComp (75.9%). Even more remarkable is that GLM-5 was trained entirely on China’s domestic Huawei Ascend chips despite, or maybe in spite, of U.S. export restrictions on advanced Nvidia hardware.
Opus Benchmark Results
Tirias Research
Zhipu, by design, also disruptively released GLM-5 under an MIT license. This allows anyone worldwide to download, modify and deploy the model at very minimal cost. By making near-frontier intelligence a commodity, they are applying direct pressure to incumbents to rethink pricing and openness.
On March 16, Zhipu released GLM-5-Turbo, a closed-source, API-only variant of its flagship GLM-5. Turbo is specially enhanced for long-running OpenClaw-style workflows. Zhipu did this, recognizing that the real value in AI is shifting from chat interfaces to autonomous agents that plan, use tools and execute long-horizon tasks without supervision. Investor enthusiasm sent its shares surging as much as 16% in a single day.
GLM-5-Turbo is not open-source licensed. It’s a proprietary, faster, cheaper API variant (around $1.20 per million input tokens, which is a fraction of Western rivals). Turbo has been fine-tuned for “OpenClaw scenarios,” which require persistent execution over hours and days, precise tool calling, complex instruction decomposition and low error rates in multi-step chains. These are the exact frustrations that cause general-purpose models to falter, but where reliable agents can deliver real productivity gains. In tasks like handling emails and calendars, automating workflows and running enterprise pipelines, these traits are crucial.
The viral adoption of OpenClaw in China has set off a self-reinforcing cycle of data collection and adoption. Tencent and other tech giants are hosting OpenClaw installation events that have drawn thousands of attendees. Local governments, including Shenzhen’s Longgang district, are pouring millions of yuan into startups and individuals building on the framework. Ecosystems are being created around it, even as Beijing is raising serious security concerns over the risks posed by open-source AI agents.
This grassroots surge in China is far ahead of anywhere else in the world. This means millions of daily agent interactions producing a vast set of high-quality data that capture real-world tool calls, task breakdowns, failures, successes and retries. By optimizing and integrating Turbo deeply into this new framework, including with its new “OpenClaw Packages,” Zhipu is positioning itself to capture the lion’s share of that data. The company is promising to feed what Turbo learns back into future open-source GLM releases. This will drive better agent performance, causing wider adoption, which will generate even richer data, which will power the next generation of models.
Execution quality isn’t a nice-to-have in agentic AI; it’s foundational. One misplaced tool call or repeated failed retries can collapse an entire workflow. China’s rapidly expanding “agent economy,” where OpenClaw has become the de facto infrastructure, will give Zhipu a built-in advantage through its targeted optimization of Turbo. Western players, like Anthropic and OpenAI field powerful models, but they have nothing comparable in scale to the real-world framework-specific data and iterative feedback loop to which Zhipu has access.
Giving away technology for free to build a massive and loyal following, then charging for the premium version is a proven strategy for technology companies. By applying it to frontier AI, it will help Zhipu beat rivals on price and performance, all the while funding serious research.
At present, agentic AI is the next big thing, and China’s grassroots enthusiasm, government backing and open-source momentum could make its labs serious contenders. Once limited by chip restrictions and thought of as an underdog, Zhipu is becoming a frontrunner in global AI. It may have figured out how to turn viral buzz into a formidable and lasting advantage.
With over 65,000 forks on GitHub, OpenClaw has remarkable community engagement. This has driven it to mature rapidly, and it has now grown into a far more hardened platform than when Tirias Research first began using it in early February. Its usability, security and maturity continues to improve with its transition to being foundation-backed, and with players like Nvidia adding privacy and security controls via its NemoClaw stack.
For anyone paying attention, including investors and developers, the message being delivered is simple. The age of AI agents is arriving, and the biggest edge may not be in having the most powerful model. The bigger edge may be in having access to the largest and most active communities that are putting those agents to work. And right now, the Chinese are “raising lobsters” at a scale the rest of the world may struggle to match.
Tirias Research tracks and consults for companies throughout the electronics ecosystem from semiconductors to systems and sensors to the cloud. Members of the Tirias Research team have consulted for IBM, Nvidia, Qualcomm, AMD and other companies throughout the data center, AI and Quantum ecosystems.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。