2025年8月26日,智谱正式上线并面向全球开源原生多模态模型 GLM-5.3-Flash(320B-A18B),同时开放 API 调用 1。匿名模型 Ox-Alpha 在 OpenCode 和 OpenRouter 上创下双平台调用量新高 1。在 Artificial Analysis Intelligence Index(AA 综合智能指数)评测中,GLM-5.3-Flash 取得 57 分,与 Claude Opus 4.8 得分持平 1。
在技术架构上,该模型总参数量为 320B(3200 亿),激活参数为 18B(180 亿),层数从 92 降至 45,并结合最新的 30T Token 多模态预训练语料 1。在定价方面,GLM-5.3-Flash 定价为 GLM-5.3 的 1/10,限时折扣内为 GLM-5.3 的 1/20,为 Opus 4.8 的 1/40 1。此外,智谱推出的 GLM Coding Plan 每天限量发放 10,000 张体验卡 1。
On August 26, 2025, Zhipu launched and open-sourced the GLM-5.3-Flash (320B-A18B) native multimodal model globally, making it available for API calls 1. The model features a total of 320 billion parameters, with only 18 billion active parameters 1. Operating under the anonymous name Ox-Alpha, the model has set new records for call volume on both the OpenCode and OpenRouter platforms 1. In terms of performance, it scored 57 on the Artificial Analysis Intelligence Index, tying with Claude Opus 4.8 1.
Architecturally, the model's layer count has been reduced from 92 to 45, and it incorporates the latest 30 trillion tokens of multimodal pre-training corpus 1. Regarding pricing, GLM-5.3-Flash is normally priced at one-tenth of GLM-5.3, but during a limited-time discount, it costs one-twentieth of GLM-5.3 and one-fortieth of Opus 4.8 1. Furthermore, the GLM Coding Plan is issuing a limited daily quota of 10,000 experience cards 1.
评论
还没有评论,欢迎留下第一条。