腾讯混元发布 224B 参数新模型,价格仅为 770B 版本的零头
To create a game: https://t.co/IsTLdr2InA
腾讯混元新模型只有 12B 激活参数,却能打过自家 770B 的 Hy4 preview,输入价格才 0.07 美元每百万 token。预算敏感的团队可以直接对比这个和 DeepSeek 的报价,性价比路线又多了一个选择。
腾讯混元(Tencent Hunyuan)发布一款总参数 224B、激活参数 12B 的新模型,输入 token 定价为每百万 0.
07 美元。该模型在性能上略超腾讯自家 770B 参数的 Hy4 preview 版本,而价格只是后者的一小部分。
这得益于 MoE 稀疏激活架构:每次推理只调用 12B 参数参与计算,在大幅降低算力和成本的同时保住大模型的容量上限。从行业角度看,以小激活参数对标大稠密模型的做法与 DeepSeek 等中国厂商的低价路线一致,说明 MoE 已成为控制推理成本的主流方案。
To create a game: https://t.co/IsTLdr2InA
To create a game: x.com/sean_wallace_/… Sean Wallace @sean_wallace_ Tencent has just released a new model featuring 224 billion (224B) parameters, with 12 billion (12B) active at any given time. It's priced at just $0.07 per million (1M) input tokens. Even more remarkably, it slightly outperforms Tencent’s own 770-billion (770B) parameter model, the Hy4 preview version, while costing only a fraction of the price. How is this possible? Let’s take a closer look. 🧵 @TencentHunyuan @WorkBuddy_AI Use WorkBuddy for free until November 30. Give it a try at no cost!
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Tencent, Hunyuan, MoE, open-source models.
Source attribution: curated from Hunyuan (@TXhunyuan).