Realtime AI News
Zhipu AI Releases GLM-5.2 Open-Source Model Alongside FP8 Quantized Variant
Zhipu AI has officially published the GLM-5.2 base model and its FP8 quantized version GLM-5.2-FP8 on Hugging Face. Both are text-generation conversational models built on a Mixture-of-Experts architecture, supporting Chinese and English with strong early community traction.
On July 2, Zhipu AI (智谱AI) released the GLM-5.2 series on Hugging Face, consisting of the base GLM-5.2 model and the quantized GLM-5.2-FP8 variant. This marks a significant open-source model update from one of China's leading AI labs.
GLM-5.2 is a text-generation conversational model built on the transformers library, employing a Mixture-of-Experts architecture via glm_moe_dsa. It supports both Chinese and English, and the accompanying technical paper is available on arXiv (reference 2602.15763).
The GLM-5.2-FP8 variant, published simultaneously, applies FP8 quantization to the base model. FP8 quantization preserves most of the model's generation quality while substantially reducing GPU memory footprint and inference latency, making it well-suited for cost-sensitive production deployments.
Community reception has been strong. On Hugging Face, the base GLM-5.2 has amassed over 3,197 likes and 176,000 downloads, while the FP8 version has surpassed 1.05 million downloads — a clear signal that developers are prioritizing efficient, low-cost deployment options.
The GLM family has consistently been a pillar of China's open-source LLM ecosystem. GLM-5.2 continues this tradition with its MoE architecture and bilingual capabilities, positioning it as a relevant option for Chinese-language AI applications.
The remarkably higher download count for the FP8 variant suggests that the developer community is actively seeking quantized models suitable for production serving. This trend may influence how other model providers structure their release strategies going forward.
Observers will now watch for GLM-5.2 benchmark results across standard evaluations, and whether Zhipu plans to release larger-scale versions or domain-specific fine-tuned variants. The arXiv paper will provide deeper technical insight for the research community.
Sources
Why it matters
Zhipu's GLM-5.2 release strengthens China's open-source LLM contributions, and the FP8 variant's outsized download volume validates strong market demand for efficient quantized deployment, likely influencing release strategies across the model ecosystem.