Chinese open model GLM-5.2 stirs Silicon Valley

TL;DR:

  • z.AI’s new open-source model GLM-5.2 is generating the strongest China-AI buzz since DeepSeek’s R1 over a year ago.
  • Built for long coding tasks and agentic workflows, it claims a one-million-token context window comparable to leading US models.
  • Being open-source, it can be downloaded and run on a company’s own systems — a draw for cost-conscious firms.

A new Chinese model is once again rattling assumptions about America’s grip on frontier AI. GLM-5.2, an open-source system from z.AI aimed at long coding tasks and agentic workflows, has drawn effusive praise across Silicon Valley — with Vercel chief executive Guillermo Rauch saying he was “almost shocked” at how good it is at coding, and a former Meta and Google DeepMind executive calling it the first open model good enough to be a “daily driver”.

The open-source challenge, again

The company says GLM-5.2 runs on a one-million-token context window, putting it in the same bracket as Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5. Its significance, though, lies less in benchmarks than in licensing. Because it is open, anyone can download it, run it inside their own systems and modify it — unlike the closed models from OpenAI and Anthropic, where customers depend on the provider. If an open model performs as well or better, it can capture market share that closed providers had assumed was theirs.

The pattern rhymes with January 2024, when DeepSeek’s low-cost R1 forced investors to question whether the US lead was as secure as it looked. Anthropic has since warned that China is narrowing the gap through looser chip controls and “distillation attacks”, arguing the US and its allies retain a chance to lock in a 12-to-24-month edge — but not indefinitely. The export-control fights of recent weeks, including the brief block on foreign access to Anthropic’s models, are the policy expression of exactly that anxiety.

Looking forward

For UK businesses, the rise of capable open models is more opportunity than threat. Self-hostable systems sidestep both subscription costs and the geopolitical risk of access being switched off in Washington — a calculation already driving UK interest in open-source and sovereign AI. The trade-offs around security, support and provenance are real, but GLM-5.2 is another data point suggesting the frontier is no longer the exclusive preserve of a handful of US labs.

Share this article