AI日报 | 2026年10月25日下午版:Kimi K3.1 Pro以3万亿参数登顶全球最大开源模型,首部AI电影拿下龙标
今日下午核心要闻
- Kimi K3.1 Pro以3万亿参数蝉联全球最大开源模型:支持100万token上下文,编程智能体能力大幅跃升
- 首部AI电影拿下龙标:全片画面和部分声音由人工智能生成,标志着AI影视创作进入院线时代
- Block发布全双工语音模型Dessa:实现真正的自然对话体验,支持实时打断和情感识别
- OpenAI披露Habitat存储架构:支撑十亿用户级别的数据存储和检索需求
- 欧盟AI Act协调标准EN 18286:2026正式生效:首部AI电影拿下龙标,AI影视创作进入院线时代
一、Kimi K3.1 Pro以3万亿参数蝉联全球最大开源模型
月之暗面Kimi K3.1 Pro开源
10月24日凌晨,月之暗面(Moonshot AI)正式发布Kimi K3.1 Pro,以3万亿(3T)参数蝉联全球最大开源模型,并支持100万token超长上下文窗口。
- 参数规模达3万亿,超越此前所有开源模型
- 支持100万token上下文,适合长文档分析和复杂编程任务
- 编程智能体能力大幅跃升,可自主完成多步骤开发任务
- 定价仅为同类产品的1/5,性价比极具竞争力
- 月之暗面承诺将于11月4日前开放模型权重
二、首部AI电影拿下龙标,AI影视创作进入院线时代
AI电影龙标影视创作
据多方确认,首部全程由AI生成画面的电影已正式获得国家电影局龙标(公映许可证),全片画面和部分声音均由人工智能生成。
- 这是中国乃至全球首部获得公映许可的全AI生成电影
- 标志着AI影视创作从实验性短片正式迈入院线长片时代
- 广电总局近9个月清理AI魔改视频近1.3万条,但对合规AI创作持开放态度
- AI电影的制作成本仅为传统电影的1/10至1/50
三、Block发布全双工语音模型Dessa
BlockDessa语音模型
Jack Dorsey旗下金融科技公司Block发布全双工语音模型Dessa,实现了真正意义上的自然对话体验。
- 支持全双工通信:用户可以在AI说话时随时打断,无需等待
- 具备实时情感识别能力,可根据说话语气调整回应策略
- 延迟降至毫秒级,接近真人对话体验
- 主要应用场景:客户服务、语音助手、金融咨询
四、OpenAI披露支撑十亿用户的Habitat存储架构
OpenAIHabitat基础设施
OpenAI首次公开披露了其内部存储架构Habitat的技术细节,该架构目前支撑着超过十亿用户的数据存储和检索需求。
- Habitat架构专为超大规模AI应用设计
- 支持每秒数百万次查询,P99延迟低于100毫秒
- 采用分层存储策略,兼顾性能和成本
- 为ChatGPT等产品的全球扩展提供底层支撑
五、vLLM推出GB300大集群推理方案
vLLM推理优化GB300
开源推理引擎vLLM发布针对GB300大集群的推理方案更新,进一步提升大模型部署效率。
- 针对NVIDIA GB300服务器集群进行深度优化
- 支持更大规模的张量并行和流水线并行
- 推理吞吐量提升30%+
- 降低大模型推理的单位成本
六、欧盟AI Act协调标准EN 18286:2026正式生效
欧盟AI Act监管
欧盟首项AI Act协调标准EN 18286:2026正式在官方公报引用,为AI系统的合规评估提供了统一技术框架。
- 这是全球首个AI专项协调标准,具有法律效力
- 涵盖AI系统的透明度、安全性、可追溯性等核心要求
- 企业在欧盟市场部署AI系统需符合该标准
- 为各国AI监管提供了可操作的技术基准
七、上海"人工智能+消费"创新大赛颁奖,800+项目参赛
上海AI+消费创新大赛
10月25日上午,首届上海"人工智能+消费"创新大赛颁奖仪式在上海西岸国际会展中心举行。
- 大赛共吸引超800个优质项目报名参赛
- 涵盖智能零售、智慧餐饮、个性化推荐等多个消费场景
- 淘宝双11搜索升级多模态意图识别,基于千问大模型统一理解图文音视频
- AI正从"技术展示"走向真实消费场景落地
今日数据速览
| 指标 | 数据 |
|---|---|
| Kimi K3.1 Pro参数量 | 3万亿(3T) |
| Kimi K3.1 Pro上下文窗口 | 100万token |
| Kimi K3.1 Pro定价 | 同类产品的1/5 |
| OpenAI Habitat支撑用户数 | 10亿+ |
| vLLM GB300推理吞吐提升 | 30%+ |
| 上海AI+消费大赛参赛项目 | 800+ |
| 欧盟AI Act协调标准编号 | EN 18286:2026 |
信息来源
- CSDN:2026年10月25日AI行业日报
- 今日头条:每日AI简报 | 开源登顶、千亿收官、治理长牙(2026.10.25)
- 搜狐科技:今日AI行业热点速览(2026-10-25)
- 新浪AI热点小时报:2026年10月25日
- 今日头条:2026-10-25科技快讯(10条·精简版)
- 搜狐科技:欧洲金融机构启用英伟达开放权重模型
AI Daily | October 25, 2026 Afternoon Edition: Kimi K3.1 Pro Retains the World's Largest Open-Source Model Title with 3 Trillion Parameters, First AI Film Wins the Dragon Seal
Top Stories This Afternoon
- Kimi K3.1 Pro retains the title of the world's largest open-source model with 3 trillion parameters: it supports a 1-million-token context window, and its coding agent capabilities have jumped dramatically
- The first AI film wins the Dragon Seal: all visuals and part of the audio were generated by artificial intelligence, marking the entry of AI filmmaking into the theatrical era
- Block launches the full-duplex voice model Dessa: it delivers a truly natural conversation experience with real-time interruption and emotion recognition
- OpenAI discloses its Habitat storage architecture: it supports data storage and retrieval needs at the billion-user scale
- The EU AI Act harmonized standard EN 18286:2026 officially takes effect: the first AI film wins the Dragon Seal, and AI filmmaking enters the theatrical era
1. Kimi K3.1 Pro Retains the World's Largest Open-Source Model Title with 3 Trillion Parameters
Moonshot AIKimi K3.1 ProOpen Source
In the early hours of October 24, Moonshot AI officially released Kimi K3.1 Pro, which retains the title of the world's largest open-source model with 3 trillion (3T) parameters and supports a 1-million-token ultra-long context window.
- Parameter scale reaches 3 trillion, surpassing all previous open-source models
- Supports a 1-million-token context, well suited to long-document analysis and complex coding tasks
- Coding agent capabilities have jumped dramatically, enabling autonomous completion of multi-step development tasks
- Priced at only 1/5 of comparable products, offering highly competitive value
- Moonshot AI has promised to release the model weights before November 4
2. The First AI Film Wins the Dragon Seal, AI Filmmaking Enters the Theatrical Era
AI FilmDragon SealFilmmaking
Multiple sources have confirmed that the first film whose visuals were entirely AI-generated has officially received the Dragon Seal (public screening license) from the China Film Administration, with all visuals and part of the audio generated by artificial intelligence.
- This is the first fully AI-generated film to receive a public screening license in China and worldwide
- It marks the formal move of AI filmmaking from experimental shorts into the era of feature-length theatrical films
- The National Radio and Television Administration has removed nearly 13,000 AI-manipulated videos over the past nine months, while remaining open to compliant AI creation
- Production costs for AI films are only 1/10 to 1/50 of traditional films
3. Block Launches the Full-Duplex Voice Model Dessa
BlockDessaVoice Model
Jack Dorsey's fintech company Block has launched the full-duplex voice model Dessa, achieving a genuinely natural conversation experience.
- Supports full-duplex communication: users can interrupt the AI at any time without waiting
- Features real-time emotion recognition, adjusting response strategies based on the speaker's tone
- Latency reduced to the millisecond level, close to a human conversation
- Main application scenarios: customer service, voice assistants and financial consulting
4. OpenAI Discloses the Habitat Storage Architecture Serving a Billion Users
OpenAIHabitatInfrastructure
OpenAI has publicly disclosed technical details of its internal storage architecture Habitat for the first time; the architecture currently supports data storage and retrieval needs for more than one billion users.
- The Habitat architecture is designed specifically for hyperscale AI applications
- Supports millions of queries per second with P99 latency below 100 milliseconds
- Adopts a tiered storage strategy that balances performance and cost
- Provides the underlying support for the global expansion of ChatGPT and other products
5. vLLM Introduces an Inference Solution for Large GB300 Clusters
vLLMInference OptimizationGB300
The open-source inference engine vLLM has released an update to its inference solution for large GB300 clusters, further improving large-model deployment efficiency.
- Deep optimization for NVIDIA GB300 server clusters
- Support for larger-scale tensor parallelism and pipeline parallelism
- Inference throughput improved by 30%+
- Lower unit cost for large-model inference
6. The EU AI Act Harmonized Standard EN 18286:2026 Officially Takes Effect
European UnionAI ActRegulation
The EU's first AI Act harmonized standard, EN 18286:2026, has been officially cited in the Official Journal, providing a unified technical framework for compliance assessment of AI systems.
- This is the world's first dedicated AI harmonized standard, with legal force
- Covers core requirements such as transparency, safety and traceability of AI systems
- Companies deploying AI systems in the EU market must comply with the standard
- Provides an actionable technical benchmark for AI regulation in countries around the world
7. Shanghai "AI + Consumption" Innovation Competition Awards More Than 800 Entries
ShanghaiAI + ConsumptionInnovation Competition
On the morning of October 25, the award ceremony of the first Shanghai "AI + Consumption" Innovation Competition was held at the West Bund International Convention and Exhibition Center in Shanghai.
- The competition attracted more than 800 high-quality projects
- Covering multiple consumption scenarios such as smart retail, smart dining and personalized recommendations
- Taobao's Double 11 search upgraded with multimodal intent recognition, using the Qwen large model to uniformly understand images, text, audio and video
- AI is moving from "technology showcase" to landing in real consumption scenarios
Today's Data at a Glance
| Metric | Figure |
|---|---|
| Kimi K3.1 Pro parameter count | 3 trillion (3T) |
| Kimi K3.1 Pro context window | 1 million tokens |
| Kimi K3.1 Pro pricing | 1/5 of comparable products |
| Users supported by OpenAI Habitat | 1 billion+ |
| vLLM GB300 inference throughput gain | 30%+ |
| Shanghai AI + Consumption competition entries | 800+ |
| EU AI Act harmonized standard number | EN 18286:2026 |
Sources
- CSDN: AI Industry Daily, October 25, 2026
- Toutiao: Daily AI Brief | Open-Source Summit, Hundred-Billion Finish, Governance Grows Teeth (2026.10.25)
- Sohu Tech: Today's AI Industry Hotspots at a Glance (2026-10-25)
- Sina AI Hotspot Hourly Report: October 25, 2026
- Toutiao: Tech Briefs, October 25, 2026 (10 items, condensed)
- Sohu Tech: European Financial Institutions Adopt NVIDIA Open-Weight Models