MiniMax 推出 MiniMax-01 系列，這是一個新的基礎模型系列，旨在處理超長上下文並增強 AI 代理開發

如今，MiniMax 在美國可能最為人所知，因為它是 Hailuo 背後的新加坡公司，Hailuo 是一種逼真的高解析度生成 AI 視訊模型，可與 Runway、OpenAI 的 Sora 和 Luma AI 的 Dream Machine 競爭。

MiniMax, a Singaporean company, is perhaps best known in the U.S. for its realistic, high-resolution generative AI video model Hailuo, which competes with Runway, OpenAI’s Sora and Luma AI’s Dream Machine.

MiniMax 是一家新加坡公司，在美國最出名的可能是其逼真的高解析度生成 AI 視訊模型 Hailuo，該模型與 Runway、OpenAI 的 Sora 和 Luma AI 的 Dream Machine 競爭。

But MiniMax has far more tricks up its sleeve: Today, for instance, it announced the release and open-sourcing of the MiniMax-01 series, a new family of models built to handle ultra-long contexts and enhance AI agent development.

但 MiniMax 還有更多的技巧：例如，今天，它宣布發布並開源 MiniMax-01 系列，這是一個新的模型系列，旨在處理超長上下文並增強 AI 代理開發。

The series includes MiniMax-Text-01, a foundation large language model (LLM), and MiniMax-VL-01, a visual multi-modal model.

該系列包括基礎大語言模型 (LLM) MiniMax-Text-01 和視覺多模態模型 MiniMax-VL-01。

A massive context window

巨大的上下文窗口

MiniMax-Text-01, is of particular note for enabling up to 4 million tokens in its context window — the equivalent of a small library’s worth of books. The context window is how much information the LLM can handle in one input/output exchange, with words and concepts represented as numerical “tokens,” the LLM’s own internal mathematical abstraction of the data it was trained on.

MiniMax-Text-01 特別值得一提的是，它可以在其上下文視窗中啟用多達 400 萬個令牌，相當於一個小型圖書館的圖書量。上下文視窗是法學碩士在一次輸入/輸出交換中可以處理的資訊量，其中單字和概念表示為數字“標記”，這是法學碩士自己對其所訓練的數據的內部數學抽象。

And, while Google previously led the pack with its Gemini 1.5 Pro model and 2 million token context window, MiniMax remarkably doubled that.

而且，雖然 Google 之前憑藉其 Gemini 1.5 Pro 模型和 200 萬個令牌上下文視窗處於領先地位，但 MiniMax 卻將其顯著增加了一倍。

As MiniMax posted on its official X account today: “MiniMax-01 efficiently processes up to 4M tokens — 20 to 32 times the capacity of other leading models. We believe MiniMax-01 is poised to support the anticipated surge in agent-related applications in the coming year, as agents increasingly require extended context handling capabilities and sustained memory.”

正如 MiniMax 今天在其官方 X 帳戶上發布的那樣：「MiniMax-01 可有效處理多達 400 萬個代幣，是其他領先型號容量的 20 至 32 倍。我們相信，隨著代理越來越需要擴展的上下文處理功能和持續內存，MiniMax-01 已準備好支援來年預期激增的代理相關應用程式。

The models are available now for download on Hugging Face and Github under a custom MiniMax license, for users to try directly on Hailuo AI Chat (a ChatGPT/Gemini/Claude competitor), and through MiniMax’s application programming interface (API), where third-party developers can link their own unique apps to them.

這些模型現在可以在自訂的MiniMax 許可下在Hugging Face 和Github 上下載，用戶可以直接在Hailuo AI Chat（ChatGPT/Gemini/Claude 的競爭對手）上進行嘗試，並透過MiniMax 的應用程式介面（API）進行嘗試，其中第三方 -派對開發者可以將自己獨特的應用程式連結到它們。

MiniMax is offering APIs for text and multi-modal processing at competitive rates:

MiniMax 以具有競爭力的價格提供用於文字和多模式處理的 API：

For comparison, OpenAI’s GPT-40 costs $2.50 per 1 million input tokens through its API, a staggering 12.5X more expensive.

相比之下，OpenAI 的 GPT-40 透過其 API 每 100 萬個輸入代幣的成本為 2.5 美元，貴出驚人的 12.5 倍。

MiniMax has also integrated a mixture of experts (MoE) framework with 32 experts to optimize scalability. This design balances computational and memory efficiency while maintaining competitive performance on key benchmarks.

MiniMax 還整合了由 32 名專家組成的混合專家 (MoE) 框架，以優化可擴展性。這種設計平衡了運算和記憶體效率，同時在關鍵基準測試上保持了有競爭力的效能。

Striking new ground with Lightning Attention Architecture

利用閃電注意力架構開闢新天地

At the heart of MiniMax-01 is a Lightning Attention mechanism, an innovative alternative to transformer architecture.

MiniMax-01 的核心是閃電注意力機制，這是變壓器架構的創新替代方案。

This design significantly reduces computational complexity. The models consist of 456 billion parameters, with 45.9 billion activated per inference.

這種設計顯著降低了計算複雜度。這些模型由 4560 億個參數組成，每次推理激活 459 億個參數。

Unlike earlier architectures, Lightning Attention employs a mix of linear and traditional SoftMax layers, achieving near-linear complexity for long inputs. SoftMax, for those like myself who are new to the concept, are the transformation of input numerals into probabilities adding up to 1, so that the LLM can approximate which meaning of the input is likeliest.

與早期的架構不同，Lightning Attention 採用了線性和傳統 SoftMax 層的混合，實現了長輸入的近線性複雜性。對於像我這樣剛接觸這個概念的人來說，SoftMax 是將輸入數字轉換為加起來為 1 的機率，以便 LLM 可以近似輸入最有可能的含義。

MiniMax has rebuilt its training and inference frameworks to support the Lightning Attention architecture. Key improvements include:

MiniMax 重建了其訓練和推理框架以支援閃電注意力架構。主要改進包括：

These advancements make MiniMax-01 models accessible for real-world applications, while maintaining affordability.

這些進步使 MiniMax-01 模型可用於實際應用，同時保持經濟實惠。

Performance and Benchmarks

性能和基準

On mainstream text and multi-modal benchmarks, MiniMax-01 rivals top-tier models like GPT-4 and Claude-3.5, with especially strong results on long-context evaluations. Notably, MiniMax-Text-01 achieved 100% accuracy on the Needle-In-A-Haystack task with a 4-million-token context.

在主流文本和多模態基準測試中，MiniMax-01 可以與 GPT-4 和 Claude-3.5 等頂級模型相媲美，尤其是在長上下文評估上取得了強勁的結果。值得注意的是，MiniMax-Text-01 在具有 400 萬個令牌上下文的「大海撈針」任務中實現了 100% 的準確率。

The models also demonstrate minimal performance degradation as input length increases.

隨著輸入長度的增加，這些模型也表現出最小的性能下降。

MiniMax plans regular updates to expand the models’ capabilities, including code and multi-modal enhancements.

MiniMax 計劃定期更新以擴展模型的功能，包括程式碼和多模式增強。

The company views open-sourcing as a step toward building foundational AI capabilities for the evolving AI agent landscape.

該公司將開源視為為不斷發展的人工智慧代理領域建立基礎人工智慧能力的一步。

With 2025 predicted to be a transformative year for AI agents, the need for sustained memory and efficient inter-agent communication is increasing. MiniMax’s innovations are designed to meet these challenges.

預計 2025 年將是 AI 智能體變革的一年，對持續記憶和高效智能體間溝通的需求正在增加。 MiniMax 的創新旨在應對這些挑戰。

Open to collaboration

開放合作

MiniMax invites developers and researchers to explore the capabilities of MiniMax-01. Beyond open-sourcing, its team welcomes technical suggestions and collaboration inquiries at model@minimaxi.com.

MiniMax 邀請開發人員和研究人員探索 MiniMax-01 的功能。除了開源之外，其團隊還歡迎透過 model@minimaxi.com 提出技術建議和合作諮詢。

With its commitment to cost-effective and scalable AI, MiniMax positions itself as a key player in shaping the AI agent era. The MiniMax-01 series offers an exciting opportunity for developers to push the boundaries of what long-context AI can achieve.

憑藉對具有成本效益和可擴展的人工智慧的承諾，MiniMax 將自己定位為塑造人工智慧代理時代的關鍵參與者。 MiniMax-01 系列為開發人員提供了一個令人興奮的機會，以突破長上下文 AI 的極限。