英伟达在 2025 年 12 月 15 日发布 Nemotron 3 系列开放模型，并同时释出训练资料与工程工具，让这家全球顶尖晶片商更像一家模型供应商。此举也可视为对冲：OpenAI、Google、Anthropic 等正打造更强自研晶片，未来可能逐步降低对英伟达硬体的依赖。文章指出，开放模型对研究与新创很关键，而美国公司虽提供小型开放模型，但更新频率不如中国竞争者，导致中国开放模型更受欢迎的趋势。
Nemotron 3 分成三个尺寸：Nano（300 亿参数）、Super（1000 亿）、Ultra（5000 亿）；参数量大致对应能力与运行成本，最大型模型往往需要整柜昂贵硬体。英伟达采取更透明做法，公开训练资料，并提供客制化与微调工具，包括一种混合式潜在混合专家架构，主打更适合打造可在电脑或网路上采取行动的代理。公司也推出可用强化学习训练代理的函式库，以模拟奖惩来调整行为。
产业面上，过去一年美国公司更倾向保密，甚至 Meta 也暗示未来释出未必维持开源。一份 OpenRouter 报告显示，2025 年经其系统送出的 tokens 中，开放模型约占三分之一；同时 DeepSeek、阿里巴巴、Moonshot AI、Z.ai、MiniMax 等中国公司持续推出强力开放模型并公开研究细节，吸引工程师实验。这对英伟达构成风险：其晶片已成为地缘政治筹码；美国在 2025 年 12 月 8 日表示将允许对中国出口上一代最佳的 H200，而中国也推动资料中心改用国产晶片，可能使中国模型更贴合本土矽基，削弱英伟达地位。
On December 15, 2025, Nvidia released the Nemotron 3 family of cutting-edge open models, plus training data and tools, signaling a shift from “chip supplier” toward “model maker.” The timing also reads as a hedge: as OpenAI, Google, and Anthropic develop increasingly capable in-house chips, they may gradually steer workloads away from Nvidia hardware. The piece notes that open models are central for experimentation and building, and that US firms’ smaller open releases are updated less frequently than China’s, a trend that has helped Chinese open models become more popular.
Nemotron 3 comes in three sizes: Nano (30B parameters), Super (100B), and Ultra (500B). Parameter count loosely tracks capability and run-cost, with the largest models often requiring racks of expensive hardware. Nvidia is also taking a more transparent approach by releasing the training data and shipping customization/fine-tuning tooling, including a hybrid latent mixture-of-experts architecture aimed at agentic systems that can take actions on computers or the web. It is also launching libraries to train agents with reinforcement learning via simulated rewards and punishments.
Industry-wide, the past year has brought less openness from US firms, and even Meta has hinted future releases may not be open source. An OpenRouter report says open models accounted for about one-third of tokens sent through its systems in 2025, while Chinese firms (DeepSeek, Alibaba, Moonshot AI, Z.ai, MiniMax) regularly publish strong open models and detailed research. That momentum could pressure Nvidia amid geopolitics: its silicon has become a bargaining chip, the US said on December 8, 2025 it would allow exports of prior-generation H200 chips to China, and China is pushing data centers toward domestic chips—potentially aligning Chinese models more tightly with Chinese silicon and weakening Nvidia’s position.