本文来源:https://mp.weixin.qq.com/s/1_tTerT8oH3E-GWoMOItlw
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
截至今天,腾讯混元大模型 Hy3 总调用量较上一代模型(Hy2)增长已超 68 倍。延续 Hy3 preview 版本的增长态势,Hy3 展现出更高的增长斜率。发布一周后,Hy3 在 OpenRouter 全球大模型调用量总榜排名第一。
在 Agent 产品中,Hy3 也展现出更强的竞争力,WorkBuddy 自选模型的用户中,选择 Hy3 的占比达到 60%。
7 月 6 日,腾讯混元 Hy3 正式发布。相比 preview 版本,Hy3 展现出显著强于同尺寸模型且比肩(参数规模2—5倍的)旗舰模型的智能水平,总体稳定性和性价比大幅提升。Hy3已在 WorkBuddy/CodeBuddy、元宝、Marvis、ima等多个业务接入,API已在腾讯云 TokenHub 以及多个海外平台上线。
开源版本模型使用门槛也在降低,就在 7 月 14 日,腾讯混元团队针对开源社区需求,把 Hy3 的权重量化到 1bit 与 4bit,让原本只能跑在多卡集群上的模型,最少一张推理显卡、乃至内存充足的本地机器就能跑。
Source: https://mp.weixin.qq.com/s/1_tTerT8oH3E-GWoMOItlw. For copyright issues, email 13951431913#139.com or submit a comment on the right.
As of today, the total call volume for the Tencent Yuanhe model Hy3 has increased by over 68 times compared to the previous generation (Hy2). Continuing the growth trend of the Hy3 preview version, Hy3 is showing a higher growth rate. One week after release, Hy3 ranked first in the global model call volume chart on OpenRouter.
In Agent products, Hy3 also exhibits stronger competitiveness. Among users of the WorkBuddy self-selection model, the proportion choosing Hy3 has reached 60%.
On July 6th, the Tencent Yuanhe Hy3 was officially released. Compared to the preview version, Hy3 shows intelligence levels significantly stronger than models of the same size and comparable to flagship models (with 2–5 times the parameter scale), with greatly improved overall stability and cost-effectiveness. Hy3 has already been integrated into multiple businesses including WorkBuddy, CodeBuddy, Yuanbao, Marvis, and ima, and its API has been released on Tencent Cloud TokenHub and several overseas platforms.
The entry barrier for using the open-source version is also being lowered. On July 14th, in response to community needs, the Tencent Yuanhe team quantized the weights of Hy3 to 1-bit and 4-bit, allowing a model that originally could only run on multi-GPU clusters to be executed on a single inference card or even on a local machine with sufficient memory.
