Hugging Face Funds 6‑Hour Free Compute for GLM‑5.2 as Musk Praises the Model

Hugging Face announced that it will provide up to six hours of free global compute for the newly released flagship model GLM‑5.2 from Zhipu AI, marking the first time the community has funded a dedicated “VIP channel” for a Chinese open‑source model.

Elon Musk responded on X, noting that GLM‑5.2 shortens the gap to the Fable‑level models and speculated that China could reach that level by the first quarter of 2027, a timeline many overseas observers called conservative.

At the same time, Anthropic’s latest Claude Fable 5 and Claude Mythos 5 were withdrawn after a U.S. export‑control order, illustrating the tightening access rules for leading closed‑source models.

On June 17, Zhipu AI officially launched GLM‑5.2 under a permissive MIT license that allows free commercial use and delivers a stable 1 million‑token context window, a substantial leap for long‑range programming tasks.

From an architectural standpoint, GLM‑5.2 introduces the IndexShare mechanism: every four sparse‑attention layers share a single indexer, reducing per‑token computation by roughly 2.9× when processing million‑token contexts.

In benchmark evaluations, GLM‑5.2 scored 51 on the Artificial Analysis leaderboard, placing it among the top three global models and the SOTA open‑source model. On code‑centric suites such as FrontierSWE and Terminal‑Bench, its performance trails the closed‑source Claude Opus 4.8 by only 1 %–4 %.

“Zhipu’s GLM‑5.2 now exceeds Google Gemini in agent capabilities,” said US open‑source researcher Nathan Lambert.

This achievement signals that the historic “big‑three” of Google, OpenAI and Anthropic are being challenged by a Chinese entrant for the first time.

Usage data from the OpenRouter platform shows Chinese models’ share rising from 1.2 % at the end of 2023 to over 50 % today, overtaking U.S. models in total request volume, although the author notes that raw usage does not equal universal superiority.

GLM‑5.2’s first‑day deployment included full compatibility with domestic accelerators such as Huawei Ascend, Pingtouge, Moore Threads, Cambricon, Kunlun, and others, mirroring the earlier ecosystem support achieved by DeepSeek V4.

When closed‑source frontier models increasingly require eligibility checks, GLM‑5.2 demonstrates an alternative path: open‑source, widely accessible, and rebuildable by developers worldwide, marking a shift from a “catch‑up” narrative to a “head‑to‑head” competition in the global AI model arena.