Token Era, Everything Intelligent | Moore Threads 2026 Product Launch: Building the All-Scenario AI Computing Foundation 词元时代，万物智能 | 摩尔线程2026产品发布会：打造全场景AI算力基石

Strategic Positioning in the "Token Era"

Moore Threads' event, themed "Token Era, All-Intelligent," positions the company at the confluence of two major trends: the exponential growth in demand for Tokens driven by Agentic AI, and the impending explosion of All-Intelligent applications. This framing is strategic, as it moves the narrative beyond raw computational power (FLOPS) to a more application-centric metric—the Token—which is the fundamental unit of intelligence for modern AI models. By anchoring its strategy here, Moore Threads is not just selling hardware, but a foundational platform for the next wave of AI, which is characterized by autonomous agents and pervasive intelligence.

Building the Three-Tier "Intelligent Foundation"

The core of the presentation is the demonstration of a full-stack, integrated ecosystem spanning cloud, edge, and end devices. This is a classic systems-level strategy aimed at capturing value across the entire AI deployment pipeline.

• Cloud: The Backbone for Large-Scale Model Training
  At the cloud level, the KUAE (夸娥) intelligent computing cluster is the flagship product. The article highlights impressive metrics: a Model FLOPs Utilization (MFU) of 60% for dense large models and 40% for Mixture-of-Experts (MoE) models, with 90% effective training time and 95% linear scaling efficiency. These figures are crucial because they demonstrate system-level engineering competence—ability to not just build clusters, but to make them run stably and efficiently at scale. The launch of the KUAE Training Suite, which covers the full pipeline from pre-training to post-training (including Reinforcement Learning) and offers compatibility with mainstream frameworks like VeRL, is a direct move to reduce the Total Cost of Ownership (TCO) and lower the barrier for developers. This indicates a focus on building a robust, developer-friendly ecosystem around its hardware.

• End: From Digital Agents to Smart Household Hubs
  The "end" segment is where Moore Threads' narrative shifts from infrastructure to user-facing, intelligent products. The digital agent "Xiaomai" (小麦) is presented as a personality-driven AI, leveraging technologies like a "two-dimensional topology memory system" for context and memory. This moves the conversation from generic assistants to agents with specialized, persistent personas.

  The hardware embodiment of this vision is the MTT AICUBE. This device is marketed as an Agentic AI hub for households, integrating the roles of an AI Agent, AI PC, and AI NAS. By packing 60+ skills and cross-app control into a single box, Moore Threads is attempting to create a Swiss Army knife for home intelligence, simplifying the user experience while centralizing data storage and processing locally. This addresses both privacy concerns and latency issues inherent in cloud-only solutions. The MTT AIBOOK similarly targets the developer and pro-user market with a focus on multi-system compatibility (Linux, Windows virtualization, Android containers) and direct agent development tools, positioning it as a "creation tool for the intelligent agent era."

Ecosystem and Software as Strategic Moats

A recurring theme is the emphasis on the MUSA ecosystem and software readiness. The claim of "Day-0 adaptation" for major domestic large models (like DeepSeek, GLM, Qwen) and support for leading inference frameworks (SGLang, vLLM) is a critical competitive point. In the GPU market, especially against established players, software ecosystem maturity and developer mindshare are as important as hardware performance. By open-sourcing vLLM-MUSA and ensuring broad model compatibility, Moore Threads is actively lowering the migration cost for developers and seeking to create a virtuous cycle of hardware adoption fueled by software support.

The presentation of "Qu A Cloud Service" (夸娥云服务) with demos like Vibe Coding (generating apps from spoken instructions) and AIGC short video production workflows further illustrates the intent to move up the value chain. It's a demonstration of turning raw compute power into tangible, creative productivity tools, showcasing the endpoint utility of their infrastructure.

Logic and Deeper Implications

The underlying logic of this full-stack approach is vertical integration and scenario dominance. By controlling the silicon ("Yangtze" SoC), the system software (MUSA, MTT AIOS), the application layer (agents like Xiaomai, MTClaw framework), and the cloud service, Moore Threads aims to deliver optimized, end-to-end solutions for specific high

核心战略：在“词元时代”筑造算力基石

本次发布会的主题“词元时代，万物智能”，深刻揭示了摩尔线程所瞄准的行业趋势与战略核心。

• 背景洞察：Agentic AI（智能体AI）正成为新范式，其自主决策、调用工具和多步推理的特性，导致对计算单元“词元（Token）”的需求呈指数级增长。这直接驱动了对底层算力的海量、高效需求。
• 战略定位：摩尔线程不再仅仅是一家GPU公司，而是致力于成为支撑这场变革的 “智算底座” 。其核心战略是构建一个从云端到边缘再到终端的、无缝协同的全栈智能计算矩阵，以应对未来万物皆需智能的场景。

云端筑基：从训练到推理的全链路能力

云端是算力的主战场，摩尔线程通过夸娥（KUAE）智算集群，展示了其作为基石的系统级能力。

1. 训练能力：万卡级集群已落地，关键性能指标（如MFU、有效训练时长）达到国际主流水平。这意味着摩尔线程已能支撑起耗资巨大的大模型全流程训练，包括预训练、微调及复杂的强化学习后训练。
2. 生态与推理速度：一个硬件平台的成败关键在生态。摩尔线程通过 “发布即适配” 策略，迅速适配了国内主流大模型（DeepSeek、Qwen等），并在SGLang、vLLM等全球主流推理框架中获得原生支持。这极大地降低了开发者的迁移门槛，是其技术走向实用的关键一步。
3. 应用转化：夸娥云服务展示了 “算力即服务” 的模式，将强大算力直接转化为生产力。例如，Vibe Coding让自然语言生成App成为可能，预示了“全民开发”的潜力；AIGC微短剧流程则证明了其在创意产业中的价值。

终端进化：让智能体“有温度”地走入家庭

如果说云端是“大脑”，终端就是与人交互的“感官”和“躯体”。摩尔线程通过自研SoC“长江”和智能体“小麦”，将AI能力真正植入个人设备。

• 智能体“小麦”：这不仅是语音助手，而是具备情景感知、长期记忆、自主任务规划能力的伴侣。其背后的MTT AIOS操作系统和二维拓扑记忆系统，旨在实现更个性化、有情感连接的交互体验，体现了从“功能型AI”向“关系型AI”的演进。
• MTT AICUBE：作为家庭AI中枢，它集成了“智能体+AI PC+AI NAS”，是一个大胆的集成化尝试。这试图解决当前智能家居设备孤岛化的问题，目标是成为家庭数据的存储中心、计算中心和智能服务调度中心。
• MTT AIBOOK：定位 “为智能体而生” ，这表明其硬件和系统设计都围绕AI开发与调试展开。支持多系统、提供丰富的工具接口，旨在为开发者提供一个从底层到应用的完整闭环开发平台。

拓展边界：从数字世界迈向物理世界

本次发布会的一个重要信号是，摩尔线程的目光已超越纯数字领域，指向了更广阔的物理世界。

• MT Lambda平台：这是 “首个全栈具身智能仿真平台” 。具身智能是指AI拥有物理身体并与环境交互（如机器人、自动驾驶）。MT Lambda通过高保真仿真，能大幅降低实体机器人的训练成本和风险，是连接AI模型与物理世界的关键桥梁。这标志着摩尔线程的算力生态正从数据中心延伸至工厂、街道等真实场景。

生态护城河：MUSA与持续进化的意义

硬件是躯体，软件生态是灵魂。摩尔线程反复强调的MUSA生态系统，是其构建长期