MiniMax 发布大模型与海螺 AI

2023 年 8 月，MiniMax 发布了自研大模型，并同步上线 AI 对话应用海螺 AI。这家由闫俊杰创立的公司，2021 年就成立，起步比 ChatGPT 热潮还早。MiniMax 给自己的定位从一开始就很清晰：多模态与效率并重。在国产模型普遍「追参数、追规模」的年代，这家公司更像一个技术派——它想做的不只是更大的模型，而是更聪明的架构。

MiniMax 真正让行业侧目，是 2025 年的 MiniMax-01 系列。这一代模型采用了一种新的 MoE 架构，搭配 400 万级别的超长上下文窗口，还引入了线性注意力的思路来降低长序列推理成本。长上下文和高效架构的组合，让 MiniMax-01 在海外开发者社区引起不小轰动——它不是靠参数堆砌，而是靠架构设计把「长文本能力」和「成本控制」同时做到。

产品层面，MiniMax 坚持「模型 + 应用」两条腿：一边开放 API 服务开发者，一边运营海螺 AI 对话与视频生成等 C 端产品。多模态能力从文本延伸到视频生成，让海螺 AI 在国产 AI 应用里占据一席之地。技术派的底色加上产品化的落地，MiniMax 走出了和「纯聊天」「纯开源」都不同的第三条路。

MiniMax 的差异化在国产模型中很有代表性。当很多人默认「中国公司只能追平海外」时，MiniMax-01 证明中国团队也能在架构层面做创新、在某些指标上反超。这不仅是公司自身的胜利，也刷新了外界对国产模型技术能力的判断。

回看 MiniMax 的历程，它让人印象最深的是那种「不跟别人卷同一个赛道」的清醒。当市场都在比谁参数大、谁榜单高时，它选择在长上下文和架构效率上做深。2026 年的国产模型格局早已不止于参数竞赛，而 MiniMax 是最早用「新架构」这一武器，为这场竞赛打开新维度的公司之一。

In August 2023 MiniMax released its self-developed large models and launched the Hailuo AI chatbot at the same time. The company, founded by Yan Junjie, had existed since 2021—before the ChatGPT wave. MiniMax's self-positioning was clear from the start: multimodal and efficiency in equal measure. In an era when domestic models commonly chased parameters and scale, this company behaved like a technical school—it aimed not at merely bigger models but at smarter architectures.

MiniMax truly turned heads with the MiniMax-01 line in 2025. That generation used a new MoE architecture, paired with a context window in the millions of tokens, and introduced linear-attention ideas to cut the cost of long-sequence inference. The combination of long context and efficient architecture caused a stir in the overseas developer community—it wasn't parameter stacking but architecture design that delivered long-text capability and cost control at once.

On the product side, MiniMax insisted on walking on two legs: opening APIs to developers while operating consumer products like Hailuo AI for chat and video generation. Multimodal ability extended from text to video generation, giving Hailuo a place among domestic AI apps. A technical-school temperament plus productization gave MiniMax a third road distinct from "pure chat" and "pure open source."

MiniMax's differentiation is highly representative among domestic models. When many assumed "Chinese companies can only catch up overseas," MiniMax-01 proved Chinese teams can innovate at the architecture level and even lead on certain metrics. This was not just a company's victory but a reset of outsiders' judgment about domestic technical capability.

Looking back at MiniMax's journey, the most striking quality is a clear-headed refusal to fight on the same track as everyone else. While the market compared parameter sizes and leaderboard scores, it chose to go deep on long context and architectural efficiency. By 2026, the domestic model landscape is no longer merely a parameter race—and MiniMax is one of the earliest companies to open a new dimension in that race with the weapon of "new architecture."