Musk Says GLM Could Reach Fable Level by Q1 2027—ZhiPu’s Tang Argues It’s Much Sooner

Last night Hugging Face (referred to as “抱抱脸”) opened a six‑hour global free‑compute channel for ZhiPu’s newly open‑sourced flagship model GLM‑5.2, covering anyone worldwide without application.

“Maybe Q1 2027.”

Elon Musk asked on X when China would reach Fable‑level capability, noting GLM‑5.2 shortens the gap. He later replied “maybe Q1 2027.” Some netizens thought the timeline was conservative, while ZhiPu’s founder and chief scientist Tang Jie quickly countered that the wait would not be that long.

“It won’t take that long.”

His comment came just after Anthropic released its strongest models, Claude Fable 5 and Claude Mythos 5, on June 9. Both models were taken offline four days later, disappearing from users’ screens.

At the same time ZhiPu announced GLM‑5.2, emphasizing that frontier intelligence should be open, usable, and built by every developer. The model was released on June 17 under the MIT license, open‑source, and permitted free commercial use.

GLM‑5.2’s technical highlights

GLM‑5.2 is designed for long‑context tasks, supporting a stable 1 million token context. It introduces an IndexShare mechanism where every four sparse‑attention layers share the same indexer, reducing per‑token computation by roughly 2.9×.

In the Code Arena front‑end development evaluation system, which involved millions of global users, GLM‑5.2 achieved the top rank among all available models.

On the Artificial Analysis leaderboard GLM‑5.2 scored 51 points, entering the global top‑three and becoming the SOTA open‑source model. On code‑centric benchmarks such as FrontierSWE and Terminal‑Bench, the gap to the leading closed‑source model Claude Opus 4.8 narrowed to 1 %–4 %.

This marks the first time an open‑source model’s coding ability matches that of top‑tier proprietary models.

Implications for Chinese AI models

The release reflects a broader shift: Chinese open‑source models have moved from low‑cost alternatives to contenders in complex, long‑context tasks traditionally dominated by Claude, GPT and other closed models.

While the market share of Chinese models on platforms like OpenRouter grew from 1.2 % at the end of 2024 to over 50 % now, performance parity remains a separate challenge. GLM‑5.2 demonstrates that high‑end capability can be open and widely accessible.

On its launch day GLM‑5.2 achieved full compatibility with domestic accelerator platforms including Huawei Ascend, Cambricon, Kunlun, and others, mirroring the ecosystem support previously seen with DeepSeek V4.

Thus, the global LLM competition now features a Chinese model that both approaches the experience of top closed‑source systems and follows an open‑source trajectory, potentially reshaping the AI landscape.