92
MiniMax has officially open-sourced its M2.7 model and announced that it has completed first-day adaptation with multiple domestic and international chip manufacturers and inference platforms, including Huawei Ascend, Moore Thread, and NVIDIA, aiming to accelerate the development of the global AI ecosystem. M MiniMax 稀宇科技
Yesterday 664 words (about 3 min) View Source →
Sign in to highlight text and take notes as you read. Sign in now
MiniMax M2.7 在今天正式开源。我们和华为昇腾、摩尔线程、沐曦、昆仑芯、NVIDIA，以及 Together AI、Fireworks、Ollama 等海内外芯片厂商、推理平台携手，在开源首日即完成模型接入与推理适配工作，推动全球 AI 生态繁荣发展。
三周前，MiniMax M2.7 率先上线。M2.7 开启了模型的自我进化，也是我们第一个 AI 深度参与迭代自己的模型。M2.7 能够自行构建复杂 Agent Harness，并基于 Agent Teams、复杂 Skills、Tool Search tool 等能力，完成高度复杂的生产力任务。
基于其在真实的软件工程、专业办公场景的优异表现，M2.7 已成为在 Hermes Agent、OpenClaw 等智能体工具中最受好评的模型之一。来自海内外的开发者与企业客户的接入需求持续增长，模型调用量在快速提升。
> * 昇腾 AI 基础软硬件实现首日适配，基于 vllm-Ascend 推理引擎在 Atlas 800 A3、Atlas 800I A2 系列产品上为模型的推理部署提供全流程支持； > > * 摩尔线程技术团队基于 MUSA 架构，针对 M2.7 的模型特点完成深度调优，成功在 MTT S5000 上实现模型的的高性能推理； > > * 沐曦曦云 C 系列 GPU 凭借全栈自研的 MXMACA 软件栈，首日完成深度适配，实现“模型发布即算力就绪”的 Day 0 体验； > > * 昆仑芯依托自研架构，通过底层算子优化与软硬件协同加速，保障 M2.7 在平台上的稳定、高效运行表现； > > * NVIDIA 推理框架 TensorRT-LLM 为 M2.7 提供了深度适配与全面优化支持，帮助开发者和企业用户高效完成模型的部署与上线。
除产品适配外，Together AI、Fireworks、Ollama、vLLM、SGLang 和智源众智 FlagOS、魔搭等海内外开发平台与社区已同步在首日接入 MiniMax M2.7 模型并完成适配。
未来，我们将与海内外更多生态伙伴携手，在模型架构优化、应用场景落地、行业生态建设等多维度展开协作，为客户和开发者提供模型的最佳使用体验，加速全球 AI 生态繁荣。
全球开源地址：
huggingface.co/MiniMaxAI/MiniMax-M2.7
github.com/MiniMax-AI/MiniMax-M2.7
Intelligence with Everyone.
Key Quotes
> MiniMax M2.7 is officially open-sourced today. We have collaborated with domestic and international chip manufacturers and inference platforms, including Huawei Ascend, Moore Thread, MetaX, Kunlunxin, NVIDIA, as well as Together AI, Fireworks, and Ollama, to complete model access and inference adaptation work on the very first day of open-sourcing, promoting the prosperity and development of the global AI ecosystem.
> M2.7 initiates the self-evolution of the model and is also our first model where AI deeply participates in its own iteration.
> M2.7 can autonomously build complex Agent Harness and, based on capabilities like Agent Teams, complex Skills, and Tool Search tool, complete highly complex productivity tasks.
> In the future, we will collaborate with more ecosystem partners, both domestic and international, across multiple dimensions such as model architecture optimization, application scenario implementation, and industry ecosystem building, to provide customers and developers with the best user experience for the model and accelerate the prosperity of the global AI ecosystem.
Tags
MiniMax
M2.7
Model Open Source
AI Ecosystem
Inference Adaptation
Related Articles
* MiniMax M2.7: Unlocking Model Self-Evolution * 40x Training Acceleration, Breaking the 'Impossible Trinity': Decrypting the MiniMax Agent RL Architecture * MiniMax Releases MMX-CLI: A Full-Modal Command Line Tool Designed for Agents * MiniMax Agent New Year Update: Good AI Products Need to Let Tools Adapt to People * EP120 From Prompt to Agent Skills: Anthropic's Ambition and the Ultimate Abstraction of LLM Applications * Orange: The Internet is Dead, Long Live the Agent * MiniMax M2.5 Released: $1/Hour, the King of Real-World Work through a native Agent RL framework.") * Domestic Models' Spring Festival Test: Front-line Review from MiniMax, GLM, and Seedance Developers | Best Ideas in early 2026, exploring core trends such as the migration from verifiable tasks to fuzzy tasks, the efficiency revolution in video models, and the implementation of AI Agents.") * #424. Exploring MiniMax: How They Build Open-Source Models * MiniMax Office Skills: Open-Sourcing a Production-Grade Office Document Engine