Introducing the Chinese-made AI 'GLM-4.7-Flash,' which is more powerful than gpt-oss-20b



Z.ai, a China-based AI company, released a lightweight AI model called ' GLM-4.7-Flash ' that runs locally on January 19, 2026. GLM-4.7-Flash has outperformed OpenAI's '

gpt-oss-20b ' in many benchmark tests.




GLM-4.7-Flash is an AI model that uses MoE architecture , which combines multiple specialized models, with a total of 30 billion parameters and 3 billion active parameters.




Below are the various benchmark results for 'GLM-4.7-Flash (total number of parameters: 30 billion, active number of parameters: 3 billion)', 'Qwen3-30B-A3B-Thinking-2507 (total number of parameters: 30 billion, active number of parameters: 3 billion)', and 'gpt-oss-20b (total number of parameters: 21 billion, active number of parameters: 3.6 billion)'. GLM-4.7-Flash recorded the highest score in most tests.



The graph above, which measures web search performance, had an error in the BrowseComp graph. The correct graph is below, showing GLM-4.7-Flash far ahead of Qwen3-30B-A3B-Thinking-2507 and gpt-oss-20b.



GLM-4.7-Flash is developed as an open model, and the model data can be downloaded from the following link. The license is the MIT License.

zai-org/GLM-4.7-Flash · Hugging Face
https://huggingface.co/zai-org/GLM-4.7-Flash



The BF16 version of GLM-4.7-Flash released by Z.ai requires 45GB or more of VRAM to run. In response to a question asking, 'Can it run on a GeForce RTX 4090?', Z.ai advises users to wait for the release of the quantized version.




Related Posts:

in AI, Posted by log1o_hf