Z.AI's GLM 5.2 opens API, MIT weights on HF, Ollama support live

Z.AI (z.ai) released GLM 5.2, initially locked behind the GLM Coding Plan but now open for everyone. Two hours ago they opened the API and pushed MIT-licensed weights to HuggingFace, with Ollama support already live. The API pricing is $1.4 per 1M input tokens and $4.4 per 1M output tokens, matching GLM 5.1. Two thinking modes are available: High and Max, with Max burning more tokens for higher scores. The model supports a 1M context window, though real-world coherence at the far end remains untested.

Benchmarks show competitive performance: 81.0 on Terminal-Bench 2.1, 62.1 on SWE-bench Pro, and 74.4 on FrontierSWE. This puts GLM 5.2 roughly on par with Opus 4.8 (trailing by one point) and slightly ahead of GPT-5.5 (edging by one point). For an open-weights model under MIT license, this is a genuinely useful position — not because it wins everything, but because it wins enough at a price that makes routing interesting.

Use cases highlighted include extracting fields from support logs, drafting internal summaries, and first-pass code review comments — the boring 70% of daily work. The author plans to widen usage if the model survives a week of real traffic without weird failures. With open weights and Ollama support, developers can run it locally or call it through any existing gateway, making it a practical alternative for cost-conscious teams.