DeepSeek新版模型单周调用量登顶全球首位 环比暴增801%DeepSeek's new model tops global weekly API calls, surging 801% month-over-month


AI free token/AI免费积分

扫一扫下载界面新闻APP

其他途径关注界面…
DeepSeek;新版模型;单周调用量;全球首位;环比暴增;801%;AI模型;人工智能;大模型;市场表现

图片来源:视觉中国

界面新闻记者 | 宋佳楠

10月26日消息,在OpenRouter发布的最新一周全球AI大模型调用量周榜单上(10月19日至25日),拿下榜首的是DeepSeek在10月17日更新的V4.1-Flash-1017版本。该模型单周调用量达到33.6万亿Token,环比增幅高达801%,成为近一周爆发力最强的模型,也是拉动国产模型总体增量的核心动力。

公开资料表明,V4.1-Flash-1017以9月10日发布的V4.1 Flash为基底,延续552B总参数、输入激活8B、输出激活16B的非对称MoE架构,原生兼容多模态,重点优化代码、自动化Agent与视觉任务。

DeepSeek新版模型单周调用量登顶全球首位 环比暴增801%
OpenRouter单周调用量排名

榜单第二名是DeepSeek旧版V4.1 Flash,9月10日正式上线,本周周调用量18万亿Token,环比下滑45%,表明开发者群体快速向新版迁移。

近期DeepSeek动作不断,先是10月20日官宣开源自研大规模推理框架DeepInfer,这套框架正是支撑V4.1 Flash系列高吞吐服务的底层底座。10月22日,据上海证券报报道,DeepSeek拟在最新一轮融资中筹得约1000亿元人民币,腾讯和宁德时代为本轮出资最多的投资方,交易预计在将来数日内正式完成。

第三名归属MiniMax M3.1 Flash,10月19日上线,周调用量13.7万亿 Token。这款模型面向复杂智能体、电脑操作、软件开发情形,官方披露推理速度提高近三倍,调用成本下降七成,主要瞄准中小开发者批量自动化任务需要。MiniMax近期也在加速商业化落地,品牌官方旗舰店登陆电商平台,上线分层订阅套餐,面向C端和小型B端使用者拓宽收入渠道。

DeepSeek新版模型单周调用量登顶全球首位 环比暴增801%

前十榜单其余席位,还涵盖智谱AI GLM 5.3 Flash、腾讯Hy4 preview,以及OpenAI、Anthropic、英伟达的海外模型。

智谱GLM 5.3 Flash发布于8月26日,使用320B总参数、18B激活参数的稀疏+线性注意力混合架构,主打长上下文Agent与代码审查,本周周调用量13.2万亿Token,环比小幅上涨8%。智谱近期的GLM 5.3 系列接入亚马逊云科技 Amazon Bedrock,启动海外云厂商分成模式,同步推进国内多家云平台合作。

腾讯Hy4 preview面向Agent、办公生产力情形持续迭代,在榜单上保持稳定席位。海外阵营里,OpenAI、Anthropic、英伟达的旗舰模型依旧守住高端推理市面,虽然单周增速亮眼,但总体份额不及国产Flash系列。

从榜单前十可以看出,Flash轻量化版本已经成为平台流量绝对主力,重参数旗舰模型不再是开发者首选。另外,百万Token长上下文窗口,几乎成为头部新模型的标配,长文档、视频理解、持续智能体任务成为刚需。再者,大模型迭代周期在持续压缩,新版本密集上线,流量切换速度越来越快,一款热门模型的长处窗口期不断缩短。竞争集中在单位成本下的有效推理能力比拼,低价、高吞吐是撬动开发者迁移最直接的筹码。

就总体情况而言,上周全球AI大模型总调用量为188万亿Token,环比增长14.63%。其中上榜国产AI大模型周调用量93.54万亿Token,环比增长18.14%;同期美国模型周调用量22.47 万亿Token,环比增长29.44%。中国大模型在该平台的周调用量,已经连续二十六周高于美国,稳居首位。

拉长到近一个月的榜单来看,市面格局变化更加突出。一个月前,海外闭源模型在榜单前排占据更多席位,国产模型席位相对分散。近三十天里,国产厂商持续高频推送迭代版本,多个Flash新品接连冲进前十,逐步挤压海外模型前排位置。

DeepSeek新版模型单周调用量登顶全球首位 环比暴增801%
OpenRouter单月调用量排名

变化背后,是三重合力共同推动。首先是架构创新,非对称MoE、混合注意力、压缩KV缓存等思路持续压低单次推理成本,把“够用能力+低价”做成标准化机型。其次是迭代节奏,国内厂商以周为单位推送微调版本,针对性补齐联网、代码、Agent短板。最后是开发者友好策略,多数新版本兼容旧API、不涨价,极大降低迁移门槛。

有分析人士提到,大模型竞争已经告别单纯跑分竞赛,转向单位Token成本下的有效产出比拼,长尾自动化需要才是这一轮增量的核心来源。

需要说明的是,上述数据只统计了OpenRouter聚合API渠道的公开调用流量,没有覆盖厂商自有官网、独立APP、私有化部署的全部请求,但其仍是全球独立开发者和中小企业需要的直观风向标。

当前全球大模型竞争早已不只是排名之争。OpenAI、Anthropic持续依靠稳定的复杂推理能力,守住大型企业高端闭源市面的基本盘;而国内厂商找到了差异化赛道,集中资源打磨高吞吐、低成本、强多模态的Flash系列,精准抓住中小开发者、自动化批量任务这片增量蓝海。两种路线并行,形成当前全球赛道的主流趋势。

在这样的行业逻辑下,榜单上的名次起落会成为常态。今天登顶的版本,随时可能被另一款快速迭代的新品超越。只有持续迭代并平衡好能力、速度与成本,才是厂商可以长期留在牌桌上的通行证。

未经正式授权严禁转载本文,侵权必究。
硬科技

微信公众号

微博

本文来源:https://www.jiemian.com/article/15240686.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Scan to download the Jiemian News APP

Follow Jiemian through other channels…

Song Jiayang · Source: Jiemian News

Image Source: Visual China

Jiemian News Reporter | Song Jiayang

On October 26, according to the latest weekly global AI large model call volume ranking released by OpenRouter (October 19 to 25), the top spot was taken by the V4.1-Flash-1017 version updated by DeepSeek on October 17. The weekly call volume of this model reached 33.6 trillion Tokens, with a month-on-month growth rate as high as 801%, making it the most explosive model in the past week and the core driving force for the overall growth of domestic models.

Public data indicates that V4.1-Flash-1017 is based on the V4.1 Flash released on September 10. It retains the asymmetric MoE architecture with 552B total parameters, 8B input activations, and 16B output activations. It is natively multimodal-compatible, with a focus on optimizing code, automated Agents, and vision tasks.

Ranked second is the old version of DeepSeek V4.1 Flash, officially launched on September 10. Its weekly call volume this week was 18 trillion Tokens, a decrease of 45% month-on-month, indicating a rapid migration of the developer community to the new version.

DeepSeek has been making continuous moves recently. First, on October 20, it officially announced the open-sourcing of its self-developed large-scale reasoning framework, DeepInfer, which serves as the underlying foundation supporting the high-throughput services of the V4.1 Flash series. On October 22, according to a report by the Shanghai Securities News, DeepSeek plans to raise approximately 100 billion RMB in its latest funding round, with Tencent and CATL being the largest investors in this round, and the transaction is expected to be officially completed within the next few days.

The third place goes to MiniMax M3.1 Flash, launched on October 19, with a weekly call volume of 13.7 trillion Tokens. This model is designed for complex agents, computer operations, and software development scenarios. The official disclosure states that its inference speed has increased by nearly three times, and the call cost has dropped by 70%, mainly targeting the batch automation task needs of small and medium-sized developers. MiniMax has also accelerated its commercialization recently; its official brand flagship store has landed on e-commerce platforms, launching tiered subscription packages to broaden revenue channels for C-end and small B-end users.

DeepSeek's new model tops global weekly API calls, surging 801% month-over-month

The remaining spots in the top ten list also include Zhipu AI GLM 5.3 Flash, Tencent Hy4 preview, as well as overseas models from OpenAI, Anthropic, and NVIDIA.

Zhipu GLM 5.3 Flash was released on August 26, utilizing a sparse + linear attention hybrid architecture with 320B total parameters and 18B activated parameters. It focuses on long-context Agents and code review. This week, its weekly call volume was 13.2 trillion Tokens, a slight increase of 8% month-on-month. Recently, Zhipu's GLM 5.3 series was integrated into Amazon Web Services' Amazon Bedrock, launching an overseas cloud vendor revenue-sharing model, while simultaneously promoting cooperation with multiple domestic cloud platforms.

Tencent Hy4 preview continues to iterate for Agent and office productivity scenarios, maintaining a stable position on the rankings. In the overseas camp, flagship models from OpenAI, Anthropic, and NVIDIA continue to hold the high-end reasoning market. Although their weekly growth rates are impressive, their overall share is lower than that of the domestic Flash series.

From the top ten rankings, it can be seen that the lightweight Flash versions have become the absolute main force of platform traffic, while heavy-parameter flagship models are no longer the developer's first choice. Additionally, million-Token long-context windows have almost become a standard feature for top new models, with long documents, video understanding, and continuous Agent tasks becoming essential needs. Furthermore, the iteration cycle of large models is continuously shrinking; new versions are launched densely, traffic switching is becoming faster, and the window period for a hot model's advantage is constantly shortening. Competition is concentrated on the comparison of effective reasoning capabilities under unit costs; low prices and high throughput are the most direct bargaining chips for driving developer migration.

In terms of the overall situation, the total weekly call volume of global AI large models last week was 188 trillion Tokens, a month-on-month growth of 14.63%. Among them, the weekly call volume of domestic AI large models on the list was 93.54 trillion Tokens, a month-on-month growth of 18.14%; during the same period, the weekly call volume of US models was 22.47 trillion Tokens, a month-on-month growth of 29.44%. The weekly call volume of Chinese large models on this platform has been higher than that of the United States for twenty-six consecutive weeks, firmly holding the top spot.

Looking at the rankings over the past month, market pattern changes are even more prominent. A month ago, overseas closed-source models occupied more spots in the front ranks, while domestic models were relatively scattered. In the past thirty days, domestic manufacturers have continuously pushed high-frequency iterative versions, with multiple new Flash products successively breaking into the top ten, gradually squeezing the front positions of overseas models.

Behind these changes is the joint push of three forces. The first is architectural innovation; ideas such as asymmetric MoE, hybrid attention, and compressed KV cache continue to lower the cost of single inference, turning "sufficient capability + low price" into a standardized model. The second is the iteration rhythm; domestic manufacturers push fine-tuned versions on a weekly basis, specifically filling gaps in networking, code, and Agents. The last is the developer-friendly strategy; most new versions are compatible with old APIs without price increases, greatly lowering the migration threshold.

Some analysts mentioned that large model competition has moved away from pure benchmarking races and turned to a competition of effective output under unit Token costs, with long-tail automation needs being the core source of this round of growth.

It needs to be explained that the above data only counts the public call traffic of the OpenRouter aggregated API channel and does not cover all requests from manufacturers' own websites, independent APPs, and private deployments. However, it still remains an intuitive barometer for global independent developers and small and medium-sized enterprises.

Currently, global large model competition is no longer just a battle for rankings. OpenAI and Anthropic continue to rely on stable complex reasoning capabilities to maintain their basic share in the high-end closed-source market for large enterprises; meanwhile, domestic manufacturers have found a differentiated track, concentrating resources to polish the high-throughput, low-cost, and strong multimodal Flash series, precisely capturing the incremental blue ocean of small and medium-sized developers and automated batch tasks. The parallel existence of these two routes has formed the mainstream trend of the current global track.

Under such industry logic, the fluctuation of rankings on the list will become the norm. The version that tops the list today may be surpassed by another fast-iterating new product at any time. Only by continuous iteration and balancing capability, speed, and cost can manufacturers remain in the game for the long term.

Hard Tech

WeChat Official Account

Weibo

DeepSeek;新版模型;单周调用量;全球首位;环比暴增;801%;AI模型;人工智能;大模型;市场表现

DeepSeek's new model tops global weekly API calls, surging 801% month-over-month

DeepSeek's new model tops global weekly API calls, surging 801% month-over-month

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry