Key Takeaways:
- KNOWLEDGE ATLAS surged 31% to HKD1,438 after open-sourcing GLM-5.2
- GLM-5.2 supports a 1 million-token context window as a fully open model
- The move follows US restrictions that pulled Anthropic's Fable-5 offline
Back
KNOWLEDGE ATLAS surges 31% as GLM-5.2 open-sourcing challenges US AI curbs
KNOWLEDGE ATLAS (02513.HK) surged as much as 48% on June 15 after fully open-sourcing GLM-5.2, its most capable model, capitalizing on US export controls that pulled Anthropic's rival Fable-5 offline three days earlier.
"GLM-5.2 is currently our most capable open-source model, supporting a truly usable 1 million-token context window and continuing to maintain leadership in long-horizon tasks," KNOWLEDGE ATLAS said in a statement.
The stock opened 15% higher at HKD1,438 and peaked at HKD1,620, a 47.7% intraday gain, before paring to HKD1,438, up 31.1%. Turnover reached HKD1.78 billion on 1.21 million shares, making it one of the most actively traded names on the Hong Kong exchange. The company said the model would be officially open-sourced this week under the GLM Coding Plan.
The open-source release comes three days after the US government ordered Anthropic to shut down Fable-5, which had topped OpenAI's GPT-5.5 on every major benchmark including SWE-Bench Pro, where Fable-5 scored 80.3% to GPT-5.5's 58.6%. Anthropic has disputed the severity of the vulnerability cited by regulators, but Fable-5 remains unavailable outside the US, creating a vacuum that Chinese AI labs are racing to fill.
GLM-5.2 is built by Z.ai (Zhipu AI), the Beijing-based company behind the GLM series. The model supports a 1 million-token context window, matching Anthropic's Fable-5 on that metric, and is positioned as a fully open alternative to proprietary Western models now subject to export controls. The company's slogan for the release — "Frontier Intelligence Belongs to Everyone" — directly challenges the US government's rationale for restricting Anthropic's models.
The competitive environment is shifting rapidly. China's Kimi K2.7 ranks second on ErdosBench, behind only the now-pulled Fable-5, while DeepSeek V4 Pro has demonstrated competitive performance at a fraction of the cost. OpenRouter's newly launched Fusion service, which polls multiple models and synthesizes answers, found that a budget trio of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro beat both GPT-5.5 and Opus 4.8 at half the cost — suggesting the gap between frontier and open models is narrowing faster than many expected.
For investors, the question is whether KNOWLEDGE ATLAS can convert developer adoption into sustained revenue. The stock's 31% surge reflects market conviction that the open-source strategy will drive global adoption, but GLM-5.2's open-source nature means the company must monetize through enterprise services and cloud inference rather than API licensing — a model that has proven challenging for even well-funded open-source AI companies.
This article is for informational purposes only and does not constitute investment advice.
