Chinese open-source GLM-5.2 hailed as first open 'daily driver' model

GLM-5.2 from Z.ai has become the week's open-model story, drawing comparisons to the original DeepSeek moment. Matt Velloso, a former Meta, DeepMind and Microsoft VP, called it the 'first open model that passes the bar as a daily driver' — meaning developers can download it, run it on private systems, and modify it freely under an MIT license, rather than treating open weights as a downgrade from closed frontier models.

Technically, GLM-5.2 ships a 1-million-token context window, significantly improved coding capabilities with flexible 'thinking effort' levels, and an improved architecture incorporating IndexShare to reduce FLOPs. Benchmarks circulating on X added fuel: Perplexity CEO Aravind Srinivas amplified a claim that GLM gets ~80% pass rate on an internal financial benchmark, beating DeepSeek v4 and Kimi; Hugging Face's account shared strong Vending Bench results; and an HN thread (551 pts) claimed GPT-5.5 hallucinates 3x more than the MIT-licensed GLM-5.2.

Competitively, the model intensifies a narrative that open Chinese models now rival closed Western frontier systems — and the timing, amid Anthropic's export-control shutdown, sharpens arguments that open weights are the only hedge against geopolitical access cutoffs. Simon Willison noted he's eager to see fast custom-silicon providers like Groq or Cerebras run GLM-5.2. Clement Delangue framed it as China leading open-source AI in the 2024-2026 window.

The caveat: a Fox News report this week cited Booz Allen findings that Chinese models emit more vulnerable code for US-based users, raising 'sleeper agent' fears that temper enterprise enthusiasm. Watch next: independent security audits and whether US enterprises adopt GLM in production.