国产大模型与算力平台协同发展成效显著

国产大模型与算力平台正在快速迭代中形成紧密协同。智谱新模型在全球测试平台OpenRouter中获得国产芯片的快速适配，其请求流量由国产芯片提供1。月之暗面发布的Kimi K3大模型总参数量达2.8万亿1，该模型推出后不到一天内便有多家国产算力平台完成了适配1。

国产算力平台虽面临模型参数量快速增长等压力，但在多方面展现出独特优势1。芯片产业的年度迭代周期与大模型几乎每月推出新版本的节奏存在错位1，这对适配能力提出了挑战。不过，国产算力平台在数据主权、部署灵活性和成本可控方面具有优势1。国家数研院副主任陈东指出，国产算力的护城河在于"成本、数据、供应链三重可控"1。稀宇科技已与9家国产算力厂商实现适配，预计本季度完成国产算力集群建设1。

面向未来，"十五五"规划明确提出加强算力设施支撑、加快培育自主可控软硬件生态1，为国产大模型与算力平台的深度协同指明了发展方向。

China's homegrown large language models and computing platforms are accelerating their mutual adaptation and integration. Zhipu's latest model achieved rapid compatibility with domestic chips during testing on OpenRouter, while Moonshot AI's Kimi K3 model—which carries 2.8 trillion parameters—was successfully deployed across multiple domestic computing platforms within a single day of its release.1

The rapid iteration cycles create both opportunities and challenges for domestic computing infrastructure. Model development timelines have compressed dramatically, with new versions launching nearly every month, while semiconductor production cycles operate on a yearly basis, creating a structural misalignment.1 Despite these hurdles, domestic computing platforms maintain distinct competitive advantages. According to Chen Dong, vice director of the National Data Research Institute, the competitive moat for domestic computing power lies in "three-fold controllability: cost, data sovereignty, and supply chain resilience."1

Industry players are moving quickly to capitalize on this convergence. Xiuyu Technologies has already completed adaptation work with nine domestic computing hardware providers and expects to finish building its domestic computing cluster infrastructure this quarter.1 National planning initiatives explicitly emphasize strengthening computing infrastructure and accelerating development of self-reliant software and hardware ecosystems, signaling sustained policy support for this domestic technology stack.1