GLM-4.7

GLM-4.7 is Zhipu AI's flagship open-weight model, released December 22, 2025 — weeks before the company's planned IPO. It's a 358-billion-parameter mixture-of-experts model with roughly 32B parameters active per token, released under the MIT license on Hugging Face. Zhipu positioned it as the top open-weight contender for coding, tool use, and multi-step agentic tasks: it ranked #1 among open-source models on Code Arena's blind coding tests (ahead of GPT-5.2 in some comparisons), scored 42.8% on Humanity's Last Exam (a 41% jump over GLM-4.6), and hit 73.8% on SWE-bench Verified. It ships with 'Reserved Thinking' and 'Turn-level Thinking' modes designed for stable, controllable long agentic rollouts.

The honest picture: some headline numbers are vendor-reported and await independent reproduction — the usual caveat for Chinese lab releases. 358B total parameters means serious hardware to self-host (this is a datacenter model, not a laptop one). Zhipu's aggressive pricing ($0.07/M input tokens) looks like a land-grab ahead of the IPO, so long-term price stability is an open question. And demand famously overwhelmed Zhipu's API capacity at launch, with new Coding Plan subscriptions rationed — the classic DeepSeek-style success problem.