科大讯飞星火X2.6国产算力

10月22日，在 2026 科大讯飞全球 1024 开发者节主论坛上，端云一体讯飞星火 X2.6 正式发布。这是首次基于全国产算力完成训练升级的大模型，标志着国产 AI 基础设施的重要突破。

核心技术亮点：

• 针对国产芯片显存与带宽限制，通过长程任务动态并行训练技术，按任务长度自适应分配计算资源，超长上下文训练效率提升约 4 倍
• 基于国产集群的自适应存算协同推理技术，结合首个基于全国产设备的四层 KVCache 技术和分布式推理技术，推理效率提升 4.5 倍
• 端侧模型星火 X2.6-4B 同步升级，未来 60% 应用可在本地完成，30% 通过云端协同完成

科大讯飞董事长刘庆峰在现场表示："大模型的安全性和极致性价比，一定要通过云端一体化来解决。"讯飞星火 X2.6 系列模型已在讯飞星火官网开放试用。

OpenAI使用政策AI治理

10月22日，OpenAI 宣布更新使用政策，新规将于 2026 年 11 月 20 日生效，主要变化包括：

• 模型路由器/AI 网关须审批：运营模型访问中介平台或让第三方直接访问 OpenAI 模型的服务，须事先获得 OpenAI 审核批准（仅在应用中集成 OpenAI 模型实现功能的不受此限）
• 禁止用输出训练竞品模型：明确禁止使用 OpenAI 模型或其输出为替代或竞争模型准备、筛选或提供训练数据及信号，包括构建训练任务、强化学习环境，以及提供打分、排序等反馈
• 推出 Safety IDs：用户可在调查过程中提供安全识别码，帮助区分不同终端用户
• 支持加州 SB53 法案强化：OpenAI 致信加州州长，支持要求 AI 企业为模型设置关闭机制、接受强制性年度独立审计

智谱医疗AI政企订单

厦门市健康医疗大数据中心"人工智能医疗大模型底座与应用工程一期"项目公布中标结果，智谱以 1058 万元中标（项目预算 1080 万元）。建设内容包括：

• 医疗大模型基础支撑平台（模型统一管理、服务调用、训练测评、算力调度）
• 数据安全协同与联邦学习能力
• 智能心理健康服务应用（覆盖筛查、导诊、干预、随访等场景）
• 数据资产管理和安全运维体系

截至 10 月 22 日，智谱及旗下公司已中标或成交至少 21 个政企项目，其中已披露金额超 1000 万元的有 3 个。智谱上半年收入约 9.54 亿元，同比增长 399.7%，已超过 2025 年全年。

MiniMax港股AI估值

10月22日，大模型企业 MiniMax（00100.HK）盘中一度跌超 7%，报 161.66 港元，跌破 165 港元的发行价，市值约 570 亿港元。该公司今年 3 月曾触及 1330 港元的历史高点，目前较高点回调超 87%。

但基本面数据并不差：

• 上半年总收入约 1.17 亿美元，同比增长 283.1%
• B 端收入同比增长 703.1% 至 7390 万美元，占比升至 63.4%
• 毛利同比增长 464.8% 至 2081 万美元
• 净亏损 3.58 亿美元，较去年同期收窄 11%

值得注意的是，10月19日晚 MiniMax 刚发布新一代 M3.1 Flash 模型，推理速度提升近 3 倍，使用成本降低 70%。股价下跌更多受港股整体科技板块下挫影响——恒生科技指数当日跌超 3%。

NVIDIAGTC柏林AI基础设施

英伟达 GTC 柏林大会（10月20-22日）虽已闭幕，但其影响仍在持续发酵。黄仁勋在主题演讲中宣布多项重磅举措：

• DSX POD 模块化 AI 工厂参考设计：部署从 6 个月缩至 14 天，2027 年初全球交付
• Vera Rubin 平台全面升级：Rubin 节点 100% 液冷，组装从 2 小时缩至 5 分钟
• 英伟达过去一年在欧洲投入近 100 亿美元，合作伙伴超 1.2 万家
• 与诺基亚在 T-Mobile 网络完成首次基于 AI 的商业电话呼叫
• 保时捷、欧莱雅、沃达丰等欧洲巨头均采用英伟达 AI 技术

黄仁勋直言："计算机产业正处于一场新工业革命的起点。"他痛批德国 AI 掉队，并呼吁欧洲应从"AI 消费者"转向"AI 生产者"。

宇树科技具身智能物理AI

在 10月22日举行的 2026 世界物理 AI 大会暨宇树合作伙伴生态大会上，宇树科技董事长王兴兴给出了一个激进的时间表：具身智能的 ChatGPT 时刻最快明年就可能实现。

他指出，对顶尖大模型公司而言，内部已基本实现模型自进化——模型自己写程序、采集数据，人只需适当引导。但从技术突破到扩散至每个人使用，仍需 5-10 年时间。

王兴兴同时警告，物理 AI 目前已看到"被击穿"的苗头，核心挑战在于 AI 模型的输入输出与真实机器人的对齐。商业落地将先于家庭场景，工业领域或成为具身智能的第一战场。

iFlytekSpark X2.6Domestic Compute

On October 22, at the main forum of the 2026 iFlytek Global 1024 Developer Festival, the end-cloud integrated iFlytek Spark X2.6 was officially released. This is the first large model to complete a training upgrade on fully domestic compute, marking an important breakthrough in domestic AI infrastructure.

Core technical highlights:

• To address the memory and bandwidth limits of domestic chips, it uses long-horizon task dynamic parallel training to allocate compute adaptively by task length, improving ultra-long-context training efficiency by about 4x
• Built on domestic clusters, its adaptive compute-storage collaborative inference technology, combined with the first four-layer KVCache technology on fully domestic equipment and distributed inference, raises inference efficiency by 4.5x
• The on-device model Spark X2.6-4B is upgraded in parallel; in the future 60% of applications can run locally and 30% through cloud collaboration

iFlytek Chairman Liu Qingfeng said on stage: "The safety and extreme cost-effectiveness of large models must be solved through cloud-device integration." The Spark X2.6 series models are now available for trial on the official iFlytek Spark website.

OpenAIUsage PolicyAI Governance

On October 22, OpenAI announced an update to its usage policy, effective November 20, 2026. The main changes include:

• Model routers and AI gateways must be approved: operating model access intermediary platforms, or services that let third parties directly access OpenAI models, requires prior review and approval by OpenAI (apps that integrate OpenAI models to deliver features are not subject to this restriction)
• Outputs may not be used to train competing models: it is explicitly prohibited to use OpenAI models or their outputs to prepare, select or provide training data and signals for substitute or competing models, including building training tasks, reinforcement learning environments, and providing scoring, ranking and other feedback
• Safety IDs introduced: users can provide a safety identifier during an investigation to help distinguish different end users
• Support for strengthening California SB53: OpenAI wrote to the California governor supporting requirements that AI companies provide a shutdown mechanism for models and accept mandatory annual independent audits

ZhipuMedical AIGovernment Contracts

The Xiamen Health and Medical Big Data Center announced the winning bid for its "Artificial Intelligence Medical Large Model Foundation and Application Project Phase I". Zhipu won with 10.58 million yuan (project budget 10.8 million yuan). The scope of work includes:

• A medical large-model foundational support platform (unified model management, service invocation, training evaluation and compute scheduling)
• Data security collaboration and federated learning capabilities
• Intelligent mental health service applications (covering screening, triage, intervention, follow-up and other scenarios)
• Data asset management and security operations systems

As of October 22, Zhipu and its subsidiaries have won or closed at least 21 government and enterprise projects, three of which have disclosed amounts above 10 million yuan. Zhipu's first-half revenue was about 954 million yuan, up 399.7% year over year and already exceeding all of 2025.

MiniMaxHong Kong StocksAI Valuation

On October 22, large-model company MiniMax (00100.HK) fell as much as 7% intraday to HK$161.66, dropping below its IPO price of HK$165, with a market value of about HK$57 billion. The company hit an all-time high of HK$1,330 in March this year and is now down more than 87% from that peak.

But the fundamentals are not bad:

• First-half total revenue of about $117 million, up 283.1% year over year
• B2B revenue up 703.1% year over year to $73.9 million, rising to 63.4% of the total
• Gross profit up 464.8% year over year to $20.81 million
• Net loss of $358 million, narrowed by 11% from a year earlier

Notably, on the evening of October 19 MiniMax had just released its new-generation M3.1 Flash model, with inference speed nearly 3x faster and usage costs down 70%. The share price decline was driven more by a broad selloff in Hong Kong tech—the Hang Seng Tech Index fell more than 3% that day.

NVIDIAGTC BerlinAI Infrastructure

Although the NVIDIA GTC Berlin conference (October 20-22) has closed, its impact continues to reverberate. In his keynote, Jensen Huang announced several major initiatives:

• DSX POD modular AI factory reference design: deployment cut from 6 months to 14 days, with global delivery in early 2027
• Vera Rubin platform fully upgraded: Rubin nodes are 100% liquid-cooled, and assembly is cut from 2 hours to 5 minutes
• NVIDIA has invested nearly $10 billion in Europe over the past year, with more than 12,000 partners
• With Nokia, it completed the first commercial AI-based phone call on T-Mobile's network
• European giants such as Porsche, L'Oréal and Vodafone have all adopted NVIDIA AI technology

Jensen Huang said bluntly: "The computer industry is at the starting point of a new industrial revolution." He sharply criticized Germany for falling behind in AI and called on Europe to move from an "AI consumer" to an "AI producer".

Unitree RoboticsEmbodied IntelligencePhysical AI

At the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference held on October 22, Unitree Robotics Chairman Wang Xingxing offered an aggressive timeline: the ChatGPT moment for embodied intelligence could arrive as early as next year.

He noted that for top large-model companies, internal model self-evolution has basically been achieved—models write their own programs and collect data, with humans only needing to guide appropriately. But going from a technical breakthrough to diffusion to everyone's use still takes 5-10 years.

Wang Xingxing also warned that physical AI is already showing signs of being "broken through", and the core challenge lies in aligning the inputs and outputs of AI models with real robots. Commercial deployment will precede home scenarios, and industry may become the first battleground for embodied intelligence.