返回资讯中心
外部精选
出海观察
#全球市场#中国出海#Heavy Hitters#AI#Highlight#News

腾讯混元Hy4 flash上线首日处理超50万亿token

来源标题:Tencent says Hunyuan Hy4 flash surpasses 50 trillion tokens in first-day usage

腾讯混元Hy4 flash于11月16日上线,首日处理超50万亿token,为混元系列首日使用量最高。该模型在OpenRouter 11月17日榜单排名第一,在WorkBuddy平台首日使用量达前代Hy4 preview的10倍,并完成超5000万真实用户任务。其总参数2240亿,推理时激活120亿参数,支持多模态和100万token上下文。

动点科技 TechNodeJessie Wu2026/11/19 · 今天9 分钟阅读中文

以下正文同步自 动点科技 TechNode,版权归原站所有,已转换为易读排版。

The rise of lightweight AI models is challenging the notion that bigger is always better. Tencent’s Hunyuan said on Wednesday that Hy4 flash processed more than 50 trillion tokens on its first day after launching on November 16, marking the highest first-day usage among models in the Hunyuan series. Hy4 flash’s strong debut was also reflected in its performance across different platforms. On OpenRouter’s daily LLM leaderboard, the model ranked first on November 17. Tencent also said that on its WorkBuddy platform, Hy4 flash recorded 10 times the first-day usage of its predecessor, Hy4 preview, and had helped complete more than 50 million real-world user tasks. images created by ChatGPT According to Tencent, Hy4 flash uses a hybrid activation architecture and has 224 billion parameters in total, including 64 billion Engram parameters, while activating 12 billion parameters during inference. The model also supports multimodal processing and a context window of up to 1 million tokens. Tencent said its generation speed reaches around 200 tokens per second. But what makes Hy4 flash particularly noteworthy is how it balances performance and cost. According to Tencent’s pricing, the model charges just 0.5 yuan ($0.07) per million input tokens and 2 yuan ($0.3) per million output tokens, with cached input tokens costing as little as 0.02 yuan ($0.003) per million. Based on Tencent’s internal WorkIndex benchmark, which evaluates common office tasks, Hy4 flash costs just one-thirteenth as much as Hy4 preview and one-twentieth as much as Hy4 to complete a single task. Its cost is also around one-eighth that of mainstream lightweight models in the same category. These advantages make it particularly suitable for high-frequency office tasks, including creating and editing presentations, document processing, spreadsheet repair, audio and video processing, and visual understanding. The launch also comes after an earlier surge in demand for Hy4. When the model debuted on WorkBuddy on November 4, demand driven by its agent capabilities exceeded expectations, with many users hitting their usage limits on the first day. Tencent subsequently expanded its Hy4 inference clusters tenfold and increased free usage quotas. This time, however, the team appears to have made more extensive preparations for high concurrency ahead of Hy4 flash’s launch, with no widespread usage-limit issues reported so far. The development also reflects a broader shift in the AI industry’s competitive landscape. As daily token usage in China continues to grow, tokens are increasingly becoming an important metric for measuring AI adoption. Meanwhile, local governments across the country are introducing new measures, ranging from computing power and model vouchers to token-based financing, to make AI services more affordable. Competition is no longer just about parameter counts or model performance. Delivering strong results while keeping operating costs low is becoming equally important. Hy4 flash’s 50-trillion-token debut sends a clear signal: the market needs not only more capable AI models, but also models that are affordable enough for frequent, everyday use. However, there is still a gap between strong first-day usage and long-term commercial value. Competition among lightweight AI models may increasingly come down to more than just pricing and token usage. The real challenge is delivering solutions to real-world business needs consistently and reliably, while keeping overall costs low. For Tencent’s Hunyuan, Hy4 flash’s strong debut is a promising start, but its ability to turn initial usage momentum into sustained user value will be a key test going forward.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →