⌘K
Change language Switch ThemeSign In
Narrow Mode
全行业都在忙着“吃虾”，MiniMax M2.7 已经让虾自己拿起筷子了
量 量子位 @克雷西
One Sentence Summary
MiniMax releases the M2.7 model, featuring SRE-level engineering capabilities and native multi-agent collaboration, while introducing a new paradigm of self-evolution where the model autonomously builds its own Agent Harness.
Summary
MiniMax has officially announced the M2.7 large model, which achieves a significant leap in reasoning, engineering capabilities, and agent collaboration. M2.7 not only matches GPT-5.3-Codex in code generation but also possesses mature SRE (Site Reliability Engineering) capabilities, capable of autonomously performing troubleshooting, log correlation, and writing security hotfix scripts. In terms of collaboration, it supports native multi-agent team formation, enabling long-process task decomposition without external frameworks. Furthermore, the model performs exceptionally well in Office automation and long-memory role-playing scenarios. The core technical breakthrough lies in its ability to self-construct and iterate on its Agent Harness, allowing it to autonomously optimize algorithm performance and train machine learning models, marking a shift for large models from 'learning to use tools' to 'autonomously creating tools' in their self-evolution phase.
Main Points
* 1. M2.7 has achieved a leap from simple code generation to SRE-level complex engineering capabilities.The model not only writes code but also correlates monitoring in production environments, performs precise troubleshooting, and writes hotfix scripts. It even understands safety red lines, such as using non-blocking syntax to avoid table locking, demonstrating extreme rigor in engineering practice. * 2. Possesses native multi-agent collaboration capabilities, enabling the formation of Agent teams without external frameworks.In scenarios involving 40 complex skills, the model maintains a 97% instruction-following rate, stably anchors identities, and makes autonomous decisions, efficiently completing the decomposition and execution of long-process complex tasks through cooperation between agents. * 3. Initiates a path of self-evolution where the model builds and iterates on its own Agent Harness.The model no longer passively uses tools provided by humans but can autonomously create tools, optimize internal code, and even optimize machine learning models in competitions, making it a truly critical link in its own R&D chain.
Metadata
AI Score
78
Website qbitai.com
Published At Today
Length 2973 words (about 12 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
##### 克雷西 发自 凹非寺
量子位 | 公众号 QbitAI
MiniMax发布M2.5仅过去一个月，再次重磅更新。
今天他们正式官宣了全新的M2.7模型，这次它干复杂任务和Agent团队协作的本事更强了。
它的推理和工程能力也有了质的飞跃，像生产线上那种让人头疼的故障排查，它自己就能搞定。
以前的模型最多也就是给你打打下手写几行代码，现在的M2.7已经是个成熟的SRE（网站可靠性工程）老手了——
自动关联监控、精准揪出Bug甚至直接写脚本把漏洞修好，它都能一条龙包揽。
最狠的是它现在能自己搭建Agent Harness，把思考和干活彻底揉在了一起，直接开启了自我进化之路。
深度适配了OpenClaw长期记忆框架之后，无论是带入真实感情陪你沉浸式玩角色扮演，还是应付那种极其复杂的Office自动化办公需求，对它来说全都不在话下。
现在，M2.7已经在MiniMax Agent和开放平台全量上线了，大家随时可以去体验一把。
最佳Cowork Agent模型
咱们先来盘一下，M2.7都有哪些硬核的亮点。
最基础的是指令遵循和多智能体协作的跃升，面对海量Skills的复杂环境，M2.7调用得极其稳健。
官方测试中，在包含40个复杂技能的场景下，它仍然可以保持97%的遵循率，并且在MM-Claw“龙虾测试”里正确率达62.7%，直逼Claude Sonnet 4.6。
另外它还原生自带多智能体协作，不用外部框架也能组建Agent Teams。模型能稳定锚定身份并自主决策，智能体之间互相配合就能把长流程任务拆解执行完。
!Image 12: img 顺着底层执行力往下，是代码能力的直观爆发。它从单纯的代码生成，拓展到了代码重构、防漏洞及复杂排障等高阶领域。
在SWE-Pro测试中，它以56.22%的正确率追平GPT-5.3-Codex，端到端项目交付的VIBE-Pro基准同样拿下高分。
拿SRE线上排障来说，一告警它就会关联监控推导原因，对调用链分析并连数据库验证，查出漏掉的索引文件后，甚至懂得先用非阻塞方式建索引止血再提代码合并。
【视频请移步公众号查看】
切换到办公场景，M2.7处理复杂Office文档同样干脆，支持Excel表格、Word文档和PPT的复杂多轮修改。
在GDPval-AA评测中，它的ELO评分位列开源第一，并超越GPT-5.3。
以真实的某公司案例为例，丢给它年报和沟通会资料，它能自主比对研报并构建营收预测模型。
分析完数据后，它能直接生成Excel透视表，顺手写出Word版调研报告，并基于模板排版出能直接拿来汇报的PPT，这已经具备了一个初级分析师的水准。
【视频请移步公众号查看】
最后在偏C端互动的角色扮演场景中，M2.7还大幅强化了人设的稳定性与对话情商。
它原生支持十种语言，跨语言交流时人格统一。深度适配长期记忆框架后，在赛博养崽等场景中展现出持久的身份认同。
基于这个特点，MiniMax官方设计并开源了一个OpenRoom交互系统，把AI了塞进万物皆可互动的Web GUI空间。
在这间赛博朋克风的“龙虾小屋”里，对话可以实时产生视觉反馈与场景交互，角色主动与环境互动，沉浸感一整个拉满。
【视频请移步公众号查看】
说了这么多，账面实力究竟有没有水分，接下来直接进入真刀真枪的实测环节。
一个模型，带动整个“AI团队”
第一个登场的是多Agent原生协作测试，但这次我们不搞什么“一人公司”，而是让一群Agent模型模拟人类玩游戏，于是直接扔给了M2.7一个“谁是卧底”的建房需求。
指令要求它不仅需要统筹1个主持人Agent和5个玩家Agent，还得给每个角色专门撰写带有独立人设的SOUL.md文件，
最后还得手搓一个后台程序和前端网页，来调度并直观呈现它们之间的游戏过程。
这个测试的刁钻之处在于，它同时深度考验了模型的指令遵循能力、统筹规划能力、多角色扮演稳定性以及全栈代码执行力。
!Image 13: img 结果，M2.7很快就把整套方案安排得明明白白，并且整理成了一目了然的格式。
!Image 14: img 当然，这个页面布局有些把“玩家”的卡片放得过大了，所以我们给了些调整要求，接下来M2.7就正式开始工作了。
一开始设计的游戏界面比较简单，但功能也是完整的，不过我们把它调整成了Minecraft风格，最后是长这个样子：
!Image 15: 图片 另一边，主持人和5个选手的“人设”，也都配置好了。
!Image 16: img 点击开始按钮后，六个原生Agent便开始基于既定规则，在这个游戏房间里进行顺畅交流。
最终，这些选手也在主持人的带领下，完全自主地跑通了整个游戏流程，并决出胜负。
【视频请移步公众号查看】
第二个测试，则是模拟真实生产环境，检验M2.7在强工程场景下的SRE级故障排查与综合推理能力。
这个任务也是放在龙虾里运行的，我们把一份“案发现场记录”直接丢在了桌面的文件夹，其中包括四份复杂生产系统文件材料。
任务要求极为苛刻，它需要像资深后端架构师那样，找出故障的直接触发事件，以及数据库CPU飙升的根本原因，并当场写出排查命令与安全的止血代码。
!Image 17: img 面对庞杂的系统材料，M2.7展现出极度敏锐的诊断直觉，迅速从乱麻般的日志中，精准锁定拖垮数据库性能的真凶，给出的EXPLAIN命令完全切中要害。
!Image 18: img 最令人放心的是，它在提供紧急恢复脚本时，极其专业地使用了PostgreSQL的CONCURRENTLY语法来执行非阻塞建库索引。
这直接证明它真真切切懂得了生产环境下“严禁锁表”的安全红线。
!Image 19: img 另外，M2.7还极为规矩地附上了完整的数据库迁移文件代码，其规范程度完全可以直接拿去提交合并请求。 整个排障流程一气呵成，彻底展现了M2.7解决复杂系统灾难的硬核实力。
大模型开始自我进化了
在M2.7这些让人眼前一亮的能力背后，还有一套技术逻辑的根本性进化。
首先一点是，M2.7已经具备了自我构建复杂Agent Harness的能力。
所谓Agent Harness，其实就是模型和现实电脑环境打交道的工具箱和操作台，OpenClaw就可以被理解为是Harness的一种。
再直白点说，以前的模型通常是人给什么工具就用什么，但现在M2.7已经能自己造工具了。
拿强化学习实验来说，研究员只要给个初始想法，M2.7就会自己把实验跑起来。
它会时刻盯着状态，自动看日志、排查故障、改代码，连提交合并请求和冒烟测试都顺手干了。
!Image 21: img 而且M2.7不仅会造工具，还能主动迭代Harness本身，给整个工具箱来一波升级。
在优化内部Agent Harness代码的测试中，M2.7全程自主把控，通过反复试错，成功摸索出了优化路径，让模型在内部评测集上的效果直接拔高了30%。
再进一步，就是模型自己去训练和升级机器学习模型，这也是它能持续提升算法性能的关键。
M2.7靠短时记忆、自反馈和自优化这三招，每次跑完一轮都会生成md文件记住教训并给下一轮指路。
把它扔进MLE Lite这种极高难度的22道竞赛题里，给足24小时让它去自主优化ML模型，最终它拿下了9金5银1铜的成绩。
眼下整个行业都在紧盯硅谷风向，跟着一起疯狂吃虾，忙着适配OpenClaw来追赶热度。
跟进开源脚手架确实能快速补齐体验，但这依然停留在教模型怎么用人类工具的层面。
但就在此时，MiniMax已经切入了下一个更重要的论题，也就是让模型自己造工具搞研发，甚至成为自身研发链条的一环。
这种主动进化的能力，正是决定下一代大模型核心竞争力的分水岭。
具备了这般能力的MiniMax M2.7，已经抢先开启了模型自我迭代的新纪元。
量 量子位 @克雷西
One Sentence Summary
MiniMax releases the M2.7 model, featuring SRE-level engineering capabilities and native multi-agent collaboration, while introducing a new paradigm of self-evolution where the model autonomously builds its own Agent Harness.
Summary
MiniMax has officially announced the M2.7 large model, which achieves a significant leap in reasoning, engineering capabilities, and agent collaboration. M2.7 not only matches GPT-5.3-Codex in code generation but also possesses mature SRE (Site Reliability Engineering) capabilities, capable of autonomously performing troubleshooting, log correlation, and writing security hotfix scripts. In terms of collaboration, it supports native multi-agent team formation, enabling long-process task decomposition without external frameworks. Furthermore, the model performs exceptionally well in Office automation and long-memory role-playing scenarios. The core technical breakthrough lies in its ability to self-construct and iterate on its Agent Harness, allowing it to autonomously optimize algorithm performance and train machine learning models, marking a shift for large models from 'learning to use tools' to 'autonomously creating tools' in their self-evolution phase.
Main Points
* 1. M2.7 has achieved a leap from simple code generation to SRE-level complex engineering capabilities.
The model not only writes code but also correlates monitoring in production environments, performs precise troubleshooting, and writes hotfix scripts. It even understands safety red lines, such as using non-blocking syntax to avoid table locking, demonstrating extreme rigor in engineering practice.
* 2. Possesses native multi-agent collaboration capabilities, enabling the formation of Agent teams without external frameworks.
In scenarios involving 40 complex skills, the model maintains a 97% instruction-following rate, stably anchors identities, and makes autonomous decisions, efficiently completing the decomposition and execution of long-process complex tasks through cooperation between agents.
* 3. Initiates a path of self-evolution where the model builds and iterates on its own Agent Harness.
The model no longer passively uses tools provided by humans but can autonomously create tools, optimize internal code, and even optimize machine learning models in competitions, making it a truly critical link in its own R&D chain.
Key Quotes
* The current M2.7 is already a seasoned SRE (Site Reliability Engineering) veteran—it can handle everything from automatically correlating monitoring and pinpointing bugs to writing scripts to fix vulnerabilities. * The most impressive part is that it can now build its own Agent Harness, completely integrating thinking and doing, directly embarking on the path of self-evolution. * This capability for active evolution is the watershed moment that will determine the core competitiveness of the next generation of large models. * Following open-source scaffolding can indeed quickly bridge the experience gap, but this still remains at the level of teaching models how to use human tools. However, at this very moment, MiniMax has cut into the next more important topic: enabling models to create their own tools for R&D.
AI Score
78
Website qbitai.com
Published At Today
Length 2973 words (about 12 min)
Tags
MiniMax M2.7
LLM
AI Agent
SRE
Code Generation
Related Articles
* 128. Manus Co-founder's Final Interview Before Sale: The Fantastical Drift to 2025... * MiniMax Hailuo Video Team's First Open-Source Release: Tokenizer Exhibits a Clear Scaling Law * Deconstructing Clawdbot: Local Architecture, Memory Management, Agent Orchestration, and Context Assembly Principles * Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * Vol.89 AI Industry 2025 Annual Summary Supplement (V4 "Can't Wait" Edition) --- 70-page PPT Solo * What You Don't Know About Claude Code: Architecture, Governance, and Engineering Practices - Tw93 * AI Starts to "Take Action", Alibaba's Qwen Leads the World * After Topping Open-Source Rankings with its Programming LLM, the Zhipu GLM Team Faced a 3-Hour Questioning Session * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * How to Design an AI Agent System HomeArticlesPodcastsVideosTweets