亚马逊在 re:Invent 发布 Nova 2 系列前沿模型，并推出 Nova Forge，允许企业在基础模型尚未完全训练完成的阶段加入自有数据进行“自定义预训练”，突破了以往只能对成品模型微调的局限。首批模型包括 Nova Lite、Nova Pro、实时语音模型 Nova Sonic，以及可处理图像、音频、视频与文本的多模态推理模型 Nova Omni。亚马逊声称 Nova 2 Pro 在多项基准中可匹敌或超越 GPT-5/5.1、Gemini Pro 2.5/3.0 和 Anthropic Sonnet 4.5，Lite 版本则与 Claude 4.5 Haiku、GPT-5 Mini、Gemini Flash 2.5 相当。Nova Forge 的策略针对 75% 美企“高度重视 AI 但缺乏专业能力与资源”这一结构性瓶颈。
Nova Forge 已被 Reddit、Booking.com、Sony、Nimbus Therapeutics 等测试。Reddit 表示传统微调无法让模型处理违规内容（因多数模型默认拒绝查看），但通过“自定义预训练 + 微调”组合，可训练出“Reddit 专家模型”，用于自动化审核。构建大型模型通常需耗资数千万至数亿美元，而基于 Nova Forge 的前沿模型成本“将大幅降低”（亚马逊未给数值）。相比开源模型缺乏训练数据透明度、而闭源模型难以深度定制，Nova Forge 提供介于两端的新路径，但模型必须在 AWS 内构建与部署，形成显著云端绑定。
亚马逊正投入数十亿美元扩建 AI 基础设施，同时用自研 Trainium 芯片支持如 Anthropic 的训练任务，提升生态竞争力。随着 OpenAI 可能成为云竞争者、谷歌与微软持续争夺客户，亚马逊借 Nova 2 系列与 Forge 提供差异化服务以巩固云市场份额。多模态 Nova Omni 的发布显示其研发能力已接近前沿，而高度可定制性被企业视为最大卖点，预示 Nova 模型在专业领域任务上可能显著优于现成通用模型。
Amazon introduced its Nova 2 frontier-model family at re:Invent and unveiled Nova Forge, a system that lets enterprises inject proprietary data during unfinished stages of model training—enabling “custom pretraining,” a capability previously limited to major AI labs. The lineup includes Nova Lite, Nova Pro, the real-time model Nova Sonic, and Nova Omni, a multi-modal reasoning system that processes images, audio, video, and text. Amazon claims Nova 2 Pro matches or surpasses GPT-5/5.1, Gemini Pro 2.5/3.0, and Anthropic Sonnet 4.5 on benchmarks, while Nova Lite performs similarly to Claude 4.5 Haiku, GPT-5 Mini, and Gemini Flash 2.5. Nova Forge directly targets the structural problem that 75% of US companies prioritize AI but lack expertise and resources to build custom models.
Early adopters include Reddit, Booking.com, Sony, and Nimbus Therapeutics. Reddit says conventional fine-tuning fails because most models refuse to analyze offensive content; combining custom pretraining with fine-tuning yielded a “Reddit expert model” suitable for automated moderation. Building frontier-scale models normally costs tens to hundreds of millions of dollars, but Amazon says Nova Forge reduces this significantly (without stating numbers). The tool occupies a middle ground between closed models (hard to customize) and open models (cheap but opaque due to undisclosed training data), though it requires customers to build and deploy models strictly within AWS.
Amazon is investing billions into AI infrastructure, using its Trainium chips to support workloads like Anthropic’s training runs as it competes with Google, Microsoft, and potentially OpenAI as a future cloud rival. Nova Omni demonstrates Amazon’s rising research strength, while customization emerges as Nova’s most attractive feature. For many real-world enterprise tasks, companies expect Nova-based custom models to outperform off-the-shelf general-purpose systems.