Models Can Also "Nest"? MiniMax Releases M2.7: The First Domestic Large Model Deeply Involved in Self-Iteration

marsbit发布于2026-03-18更新于2026-03-18

文章摘要

Artificial intelligence is evolving from monthly updates to self-evolution. On March 18, MiniMax released its first new model version deeply involved in its own iteration—MiniMax M2.7. This marks a new stage in model development: large models are no longer solely trained by human programmers but have begun to "train themselves." The core breakthrough of MiniMax M2.7 lies in its strong autonomous construction capability. It can independently build complex Agent Harness (intelligent agent testing frameworks) and, relying on underlying capabilities such as Agent Teams, complex Skills, and Tool Search tools, complete highly complex productivity tasks autonomously. In simple terms, M2.7 is not just a smarter conversational agent but also a "digital engineer" capable of self-diagnosis and self-optimization. This "self-participatory iteration" model will significantly enhance the model’s logical reasoning and tool invocation accuracy when facing unknown complex tasks. Currently, this self-evolving MiniMax M2.7 model has been fully launched on the MiniMax Agent platform and open platform. As large models begin to deeply participate in their own "growth" process, the ceiling of AI may be raised once again.

The evolution speed of artificial intelligence is transitioning from "monthly updates" to "self-evolution." On March 18, MiniMax officially released its first new version model deeply involved in iterating itself—MiniMax M2.7. This marks a new stage in model development: large models are no longer solely fed by human programmers but have begun to learn to "guide themselves."

According to reports, the core breakthrough of MiniMax M2.7 lies in its powerful autonomous construction capability. It can independently build complex Agent Harness (intelligent agent testing frameworks) and, relying on underlying capabilities such as Agent Teams (intelligent agent collaboration), complex Skills, and Tool Search tool, independently complete highly complex productivity tasks.

Simply put, M2.7 is not just a smarter conversationalist but also a "digital engineer" capable of self-diagnosis and self-optimization. This "self-participatory iteration" model will significantly enhance the model's logical reasoning limits and tool invocation accuracy when facing unknown complex tasks.

Currently, this MiniMax M2.7 model, equipped with self-evolution genes, has been fully launched on the MiniMax Agent platform and open platform. As large models begin to deeply participate in their own "growth" process, the ceiling of AI may be raised once again.

Meanwhile, the AI computing power and application market are also seeing frequent developments. LuChen Technology announced the completion of a Series B financing round worth hundreds of millions of yuan, with its overseas revenue share soaring to 79%; meanwhile, due to a surge in call volumes, some of Alibaba Cloud's AI computing power products have reportedly seen price increases. Amid the interplay of technological iteration and market fluctuations, the AI track in 2026 is becoming increasingly urgent and full of variables.

你可能也喜欢

卡尔达诺的价格反弹结束了吗？以下是ADA可能瞄准0.30美元的原因

卡尔达诺（ADA）在经历了连续三个季度的显著下跌后，于本季度初强势反弹30%，表现优于大盘。然而，其价格在触及0.20美元关键阻力位后迅速回落，显示出该位置抛压较重。同时，链上数据显示其总锁仓价值（TVL）在过去一年大幅下降近68%，表明流动性未能跟上价格走势。另一方面，积极的信号也在浮现。自去年低点以来，卡尔达诺网络新增了近1.48万个非空钱包，用户参与度在TVL下滑的背景下依然保持增长，这暗示市场信心可能正在暗中积聚。从技术面看，若能有效突破并站稳0.20美元阻力位，则有望进一步上探0.25美元，甚至迈向0.30美元的目标价位。综合来看，尽管短期面临获利了结和技术阻力，但用户基础的稳步扩大和整体山寨币市场的回暖氛围，为ADA后续走势提供了潜在支撑。市场焦点在于其能否将关键阻力转化为支撑，从而开启更持续的上涨趋势。

ambcrypto9分钟前

ambcrypto9分钟前

GPT-5.6 Sol首批内测结果来了，同任务成本只有Fable 5一半

GPT-5.6 Sol预览版首批内测结果出炉，表现强劲。英伟达工程师称其在30小时内即达到Opus需64小时的CUDA加速效果，后续版本或能碾压Opus。在图像生成（如飞船设计）和前端设计上，其细节、层次感和一致性优于GPT-5.5。该模型定位为处理高难度推理和复杂代码的长链路任务，特点是代码简洁（代码量仅为Opus的1/5）、追求深度优化而非快速试错。与竞争对手Fable 5相比，Sol在部分基准测试上持平或反超，但在整体体验和代码质量上仍略逊一筹。不过，Sol的成本优势显著：其每百万Token输入成本5美元、输出成本30美元，仅为Fable 5（输入10美元、输出50美元）的一半左右。此外，Fable 5因安全限制过严常被用户诟病，Sol的限制则相对较小。模型预计将于近期向全体用户正式开放，届时将提供更全面的实测对比。

marsbit42分钟前

marsbit42分钟前

机器人看视频学操作，伯克利首次打通互联网视频到灵巧手真机部署链路

加州大学伯克利分校的研究团队提出了一种名为“Do as I Do”的端到端流程，首次成功实现了从互联网单目RGB视频到灵巧手真机部署的完整链路。该方法旨在解决机器人“看”视频却“不会做”的核心难题，即如何将带有噪声的日常人类操作视频，转化为多指灵巧手能够执行的动作轨迹。系统流程分为两个阶段。第一阶段，通过引导式扩散模型，稳定地从视频中重建出4D手-物交互过程，解决了传统方法在运动模糊、遮挡下跟踪不稳定的问题。第二阶段，针对重建出的带噪声参考轨迹，改进了原有的动作重定向优化框架，通过多项设计显著提升了成功率，从25%提升至71%。研究团队利用该方法，从网络视频中生成了涵盖放置、书写、搅拌、锤击等20类复杂日常操作的500条已验证轨迹，并选取其中10项成功部署到双UR3e机械臂与双Sharpa Wave灵巧手平台上进行50Hz的真实执行。Sharpa Wave灵巧手因其接近人手的尺度和自由度，更适合此类迁移。研究同时指出，直接从网络海量视频中筛选可用数据至关重要，未经处理的视频中仅有约5%的片段适合用于机器人学习，并总结了一套关键的数据筛选要点。这项工作标志着将人类视频大规模转化为机器人可执行数据的关键一步，为机器人学习开辟了新的数据来源。

marsbit52分钟前

marsbit52分钟前

Robinhood Chain生态速览：哪些项目值得提前布局？

Robinhood Chain作为Robinhood进军Layer2的核心，旨在构建覆盖股票代币、RWA、永续合约和现货交易的完整生态。主网上线后同步开启了为期90天的Gas费减免活动。截至2026年7月6日，其生态总锁仓价值约为7673万美元。本文梳理了生态内值得关注的部分项目： **Arcus**：由dYdX Labs与Robinhood Crypto联合构建的DEX，专注现货和永续合约，促进股票代币的点对点交易，其未来代币将优先分配给dYdX社区成员。 **Lighter**：一个ZK驱动的去中心化永续和现货交易所，已部署在Robinhood Chain上，并为Robinhood社区提供了价值1100万美元的代币激励计划。 **Morpho**：支持Robinhood Earn功能，用户可通过自托管钱包借出稳定币USDG获取收益，资金会存入Morpho金库进行分配。 **Rialto**：一个链上现货交易所，支持加密资产、股票、ETF等多种资产，初期将上线90多种Robinhood股票代币，其核心采用propAMM（使用自有库存和定价逻辑的链上做市商）模式。 **Arrakis**：面向代币发行方的非托管链上做市商协议，帮助管理DEX上的集中流动性。 **Meridian**：专注于RWA的永续合约和预测市场平台，结算资产为USDe，其前身为Ethereal。 **Native**：一个链上价格发现和执行系统，已部署至Robinhood Chain，该项目表示目前没有代币发行计划。（请注意：本文内容不构成任何投资建议，参与者应自行研究并评估风险。）

marsbit1小时前

marsbit1小时前

想薅 Robinhood Chain 羊毛？哪些生态项目值得关注？

Robinhood Chain作为Robinhood进军Layer2的核心，主网上线后开启了为期90天的Gas费减免活动。生态TVL目前约7673万美元，生态项目包括： 1. **Arcus**：由dYdX Labs与Robinhood Crypto联合构建的DEX，专注股票代币和加密货币的现货与永续合约交易，未来代币将优先分配给dYdX社区成员。 2. **Lighter**：ZK驱动的去中心化永续和现货交易所，已投入1100万美元等值代币激励用户交易赚取积分。 3. **Morpho**：支持Robinhood Earn用户借出稳定币USDG，预期年化收益率7%，资金通过Morpho市场分配。 4. **Rialto**：链上现货交易所，支持加密资产、股票、ETF等，初期上线90多种Robinhood股票代币，采用propAMM等流动性来源。 5. **Arrakis**：非托管链上做市商协议，帮助代币发行方管理DEX流动性，曾完成400万美元种子轮融资。 6. **Meridian**：专注于RWA的永续合约和预测市场平台，前身为Ethereal，结算资产为USDe，部分治理代币将分配给ENA质押者。 7. **Native**：链上价格发现与执行系统，已部署至Robinhood Chain，目前无代币发行计划。本文不构成投资建议，请读者自行研究并评估风险。

Foresight News1小时前