← 回總覽

零跑普及世界模型智驾:8.68 万新车实现车位到车位,支持低算力平台

📅 2026-03-28 14:16 梦瑶 人工智能 14 分鐘 17172 字 評分: 88
零跑汽车 世界模型 自动驾驶 模型蒸馏 端到端
📌 一句话摘要 零跑汽车通过极致模型蒸馏技术,将具备物理因果理解能力的「世界模型」智驾方案下放到 10 万以内的入门级车型,实现了低算力平台上的高阶智驾普及。 📝 详细摘要 本文详细报道了零跑汽车在自动驾驶技术上的重大突破。零跑率先将原本属于高端豪华车型的「世界模型」技术引入 8.68 万起的入门级车型 A10,打破了高阶智驾对昂贵高算力硬件的依赖。文章深入对比了规则驱动、一段式端到端(E2E)与世界模型三种技术范式,指出世界模型通过理解物理规律和因果关系,有效解决了端到端方案在复杂场景下不可控的「黑盒」问题。在工程实现上,零跑通过「极致蒸馏」技术,成功将复杂的云端大模型部署在仅有 100
Skip to main content ![Image 14: LogoBestBlogs](https://www.bestblogs.dev/ "BestBlogs.dev")Toggle navigation menu Toggle navigation menuArticlesPodcastsVideosTweetsSourcesNewsletters

⌘K

Change language Switch ThemeSign In

Narrow Mode

8.68 万新车普及车位到车位,世界模型不吃高算力!零跑夯爆了

量子位 @梦瑶

One Sentence Summary

By leveraging extreme model distillation, Leapmotor has brought its 'World Model' autonomous driving solution—capable of understanding physical causality—to entry-level models under 100k RMB, achieving the democratization of advanced driving assistance on low-compute platforms.

Summary

This article details a major breakthrough by Leapmotor in autonomous driving technology. Leapmotor is the first to introduce 'World Model' technology, previously reserved for high-end luxury vehicles, into its entry-level A10 model starting at 86.8k RMB, breaking the reliance of advanced driving assistance on expensive, high-compute hardware. The article provides an in-depth comparison of three technical paradigms: rule-driven, one-stage end-to-end (E2E), and World Models. It highlights how World Models, by understanding physical laws and causal relationships, effectively solve the 'black box' problem of uncontrollable behavior in complex scenarios inherent in E2E solutions. Regarding engineering implementation, Leapmotor successfully deployed a complex cloud-based large model onto the Qualcomm 8650 platform with only 100 TOPS of dense computing power through 'extreme distillation' technology, demonstrating its leadership in AI engineering deployment and technological democratization.

Main Points

* 1. Leapmotor achieves 'technological democratization' by bringing World Models to entry-level vehicles under 100k RMB.By bringing the World Model solution, previously exclusive to luxury cars, to new vehicles starting at 86.8k RMB, Leapmotor has broken the cost barrier for advanced driving assistance, driving the widespread adoption of 'point-to-point' functionality. * 2. World Models outperform traditional end-to-end solutions in understanding physical laws.Unlike end-to-end behavioral cloning (a 'black box'), World Models possess cognition of 3D space, physical laws, and causal relationships. They can perform 'mental sandbox' simulations of parallel worlds, making decisions more predictable and safer. * 3. Extreme distillation technology solves the reliance of high-end AI models on massive computing power.Through knowledge distillation and neural network architecture optimization, Leapmotor successfully deployed a World Model on the Qualcomm 8650 platform with only 100 TOPS of dense computing power, achieving an engineering leap by running advanced algorithms on low-cost hardware. * 4. Leapmotor adopts a multi-path intelligent strategy to adapt to different product tiers.Applying one-stage end-to-end, VLA (Vision-Language-Action), and World Models to different vehicle models reflects their fundamental exploration of returning to the essence of the car as a mobile robot—interacting with the physical world.

Metadata

AI Score

88

Website qbitai.com

Published At Today

Length 3650 words (about 15 min)

Sign in to use highlight and note-taking features for a better reading experience. Sign in now

> 贾浩楠 发自 凹非寺 > > > 量子位 | 公众号 QbitAI

2026智能车最热黑科技——世界模型,第一个把门槛打下来的玩家,意料之外,情理之中: 零跑汽车,创造了科技“普及平权”的新纪录,四五十万豪华车的世界模型智能辅助驾驶方案,将下放到10万以内的入门级车型。

!Image 15

而且放话不只是能用,依托世界模型技术体系,AI司机从能用变成了好用。

世界模型,本身是AI模型和真实物理世界直接链接、交互,具有AGI“终局”潜力的全新范式。

而上车之后,有巨大的潜力和价值,毕竟智能汽车是眼下直通物理AI最成熟的平台——自由移动、自由定义、自由进化。

零跑入局,可能会彻底改变格局:

车端的世界模型,要从旗舰车型的零星尝试,走向普惠大众的“铺天盖地”了。

零跑世界模型智能辅助驾驶系统的实测,选在杭州浙二医院附近的核心城区,本地人一听就知道附近路况的复杂程度:

!Image 16

这也是零跑世界模型实测的第一个厉害之处:人车混行的极窄路段,敢博弈,而且轨迹策略合理,高效通过

既不鲁莽危险,也不会让用户觉得过于“胆小”。

复杂环境的多目标博弈通行,这一点是当下绝大多数量产智能辅助驾驶系统不具备的能力。

除了博弈场景,零跑世界模型在不那么复杂的路段,也展现出比正常通行更高阶的能力:

!Image 17

一个常规匝道并入的场景,仍然能看出零跑世界模型的“好用”之处:和人类司机一样,平滑规划变道时机,兼顾安全和效率。

这样无限贴近人类成熟司机,以及安全驾驶习惯的风格,还体现在和低速交通参与者的交互中:

!Image 18

整个过程没有重刹没有顿挫,也没有停车等待,而是和人类司机一样,缓慢蠕行等待行人通过。

还有一个令人印象深刻的场景:

!Image 19

路边的临停、违停指示牌,零跑世界模型不但看到了,还看懂了,聪明地选择了绕行。

所以从这几个场景体验可以看出,零跑世界模型所谓的“好用”,其实翻译成具象的表达,就是一个预期感,和你自己开车会采取的策略几乎一致,不会让你感到系统太蠢太胆小,更不会让你感到害怕。

系统对场景,从感知到认知;对决策,从模仿到思考。

就比如路边的临停、违停其实是对AI司机考验很大的科目,尤其是流行的端到端方案,大概率会把路边临停违停理解成正常等红灯的车流,尾随傻等。

反而是真正等红灯的车流,容易被标记成“慢车”,做出危险、突兀的绕行动作。

!Image 20

这是端到端的局限,本质是模仿人类司机驾驶行为,但并不理解场景,容易出现不可控的危险动作

再比如,前面说的匝道并入,目前多数量产系统都能完成,但往往从匝道汇入后,系统的规则会强制车辆线驶向最右车道,然后又在“靠左快速通行”的规则指导下再向左变道。

结果就是在高速路上“画龙”,反倒增加了风险——这是古早规则主导自动驾驶时代的“遗毒”,现在其实也不少见:

用巨量人力手写规则代码的智能辅助驾驶系统,的确也能处理绝大部分路况。但规则就像“毒药”,能快速解决难题,但写了第一条就会有第一万条,最后得到一套驾驶风格“前倨后恭”、用户体验“前俯后仰”的“屎山”算法。

!Image 21

所以量产高阶智能辅助驾驶普及之后,在2026年出现了新的竞争趋势:

各家都有车位到车位,但真正有价值的L2+产品体验,是用户“无感”的,和你的家人朋友开车一样让你安心

或者说,好的AI司机,不需要去刻意测评

朱江明在世界模型实测中,说了一句话很有意思的话:

> 如果是规则(主导),那只会先选一条固定路线。

所以可以推测,零跑世界模型目前为止是没有任何一行“if else”规则代码的。

但又和不依赖规则的端到端“黑盒”体系,有根本不同。

!Image 22

简单来说,“世界模型”首先是一个大模型,但是多模态输入:不仅通过文字,更通过视觉、触觉、声音等多种感知来理解世界。

打个比方,你扔出一个玻璃杯,ChatGPT、DeepSeek、豆包等等语言模型知道“杯子会碎”,但它并不真正理解重力、硬度、动量这些物理量是如何在时空中相互作用的。

但世界模型不一样,AI建立了对三维空间、物理规律、物体持久性、因果关系的认知,尤其是万事万物的因果关系理解。

一般世界模型架构核心由三组件构成:

> 视觉编码器/压缩器:将高维度的观测数据(如摄像头拍摄的图像、视频帧)压缩成低维度的潜在向量。 > > > 序列预测器/动态核心:使用循环神经网络(RNN) 或Transformer 架构的“物理引擎”核心 ,它接收当前状态的潜在向量和智能体(或外界)采取的动作,预测下一个状态的潜在向量。实际上是在潜在空间里“模拟”世界。 > > > 渲染器/解码器:当需要人类观察或评估模型推演的结果时,将预测出的“潜在状态”还原成可视的像素图像,它相当于“图形显卡”。

所以世界模型最大的价值,是允许“如果……会怎样”的推理式交互,理解行为与后果之间的因果链条,而不是仅仅发现统计相关性。

!Image 23

说到这你可能马上就明白了,爆火的Seedance视频生成能力,其实本质就是世界模型。

而自动驾驶的本质,就是根据实时路况,不断修正行车轨迹——世界模型完美适配需求,而且比端到端这样的轨迹输出“黑盒”更加可控。

只不过在汽车智能化浪潮中,关于世界模型存在争议:有玩家认为世界模型是单纯后端的模拟器,用于数据生成仿真训练。

但是,也有另一派玩家把车端的系统直接叫世界模型——零跑显然是这一派——车端世界模型的任务,是进行实时环境认知与决策,强调的车端“平行世界推演”能力,相当于“脑内小剧场”。

其实两派并没有根本矛盾,属于工程落地与宣传上侧重点不同,因为AI行业对于世界模型是有阶段性共识的:

> 从海量数据中自主提炼物理与社会规律,形成对世界运作方式的抽象理解。

从这个共识出发,零跑自研世界模型,首先是宏观上的,即后端的数据驱动体系,据透露目前已经建成数千卡的算力基础设施规模,并且跑通了真正的数据闭环体系。

!Image 24

意思是数据中的bug、问题,能自动走完从“被发现”到“被解决并被验证”的路径:自动发现异常行为,然后将问题自动归类、建成数据集,再针对性自动进训练/仿真,出了解决方案后,系统还能自动评估效果。

研发中“人”的作用,是定义和监管、决策,而不是自己徒手标注、调参、评估、部署……

“数据直接解决问题”是最关键的核心能力,尤其在AI大模型技术范式的革新中。

这一能力看似简单,实际对车企、自动驾驶公司来说非常稀缺,因为会直接影响泛化性、效率、成本。

而在用户一侧,零跑世界模型还是能交付、能体验的产品方案。

云端的超大参数世界模型,通过知识蒸馏的手段,再对神经网络结构,模型部署等环节做优化,实现了车端部署。

蒸馏到啥程度呢?据透露,目前测试中的零跑世界模型,依托的是高通8797芯片平台,单颗稠密算力320TOPS。

!Image 25

但这还不是极限,零跑给出的承诺是,未来所有车型中只要是搭载高通8650芯片的配置(即搭载激光雷达),都能上车世界模型。

值得注意的是,尽管官方现在没有直白说,但8.86万版型A10,实际也是激光雷达+8650配置。

也就是说……

零跑创纪录把世界模型门槛打到十万元以下,背后是另外一个技术探索的新纪录: 极致蒸馏,把世界模型部署在(稠密)算力100TOPS的算力平台

这是马斯克也从未做到过的事,但零跑也并不是违反常识和技术发展规律,董事长朱江明给出了详细解释:

> 纯粹只是把智能辅助驾驶做到好用,类似于8650的算力其实就够了。 > > > 但要说把车载AI作为大模型个人助手,就是另外一回事了。

尽管掌舵人朱江明对外曾说零跑的策略是“技术不成熟时观望,成熟就跟进”,但事实上如今的零跑在AI层面,已经超脱了单纯车企的视角和诉求。

毕竟如果只追求车位到车位的普及,一段式端到端足以胜任,而且成本会比自研世界模型更低。

另一个层面,回顾一下零跑近两年的技术进展,始终在智能化上快人一步,又不是无迹可循。

!Image 26

比如和高通合作量产了智能汽车最早的舱驾一体化落地方案,以及在中低算力平台的强功能体验,再到如今一段式端到端、VLA、世界模型多路线并进的智能化方案等等,在“性价比”之外,外界正在越来越意外地认识到一个全新的零跑。

但另一个方面,零跑又没什么变化。

仍然是技术普惠第一,把最前沿的智能化技术、体验,不设门槛普及。

事实上,在智能化普及上,零跑是车圈少见的“行胜于言”,真实交付的体验其实远超巨头车企。

从去年B10率先把激光雷达普及到10万级,就能看出来。

而刚刚上市的A10,同样是把车位到车位体验普及到8万级,行业唯一。

!Image 27

这背后,是在车企阵营中,零跑率先洞察、抓住了AI浪潮的趋势。

智能汽车、自动驾驶行业的争论,尤其是今年VLA阵营和端到端阵营的口水仗,零跑根本不站队不下场。

而是用产品方案给出了自己的理解: 一段式端到端足够成熟,给入门级产品带来全场景无断点体验; 从大语言模型借鉴而来的VLA体系,零跑放在如今的旗舰车型,给出最领先的舱驾一体智能化体验;

跳出大语言模型范式的限制回归汽车、移动机器人的本质——和物理世界交互,零跑毫不迟疑地选择去探索更底层的AI范式革新。

而世界模型一通百通,改变的又岂止是汽车?

零跑汽车2025年全年累计交付新车596,555辆,同比增长103%,并凭借这一成绩成为中国造车新势力品牌的年度销量冠军。

而在销量问鼎新势力之后,现在的零跑,开始尝试在智能化上问鼎了。

量子位 @梦瑶

One Sentence Summary

By leveraging extreme model distillation, Leapmotor has brought its 'World Model' autonomous driving solution—capable of understanding physical causality—to entry-level models under 100k RMB, achieving the democratization of advanced driving assistance on low-compute platforms.

Summary

This article details a major breakthrough by Leapmotor in autonomous driving technology. Leapmotor is the first to introduce 'World Model' technology, previously reserved for high-end luxury vehicles, into its entry-level A10 model starting at 86.8k RMB, breaking the reliance of advanced driving assistance on expensive, high-compute hardware. The article provides an in-depth comparison of three technical paradigms: rule-driven, one-stage end-to-end (E2E), and World Models. It highlights how World Models, by understanding physical laws and causal relationships, effectively solve the 'black box' problem of uncontrollable behavior in complex scenarios inherent in E2E solutions. Regarding engineering implementation, Leapmotor successfully deployed a complex cloud-based large model onto the Qualcomm 8650 platform with only 100 TOPS of dense computing power through 'extreme distillation' technology, demonstrating its leadership in AI engineering deployment and technological democratization.

Main Points

* 1. Leapmotor achieves 'technological democratization' by bringing World Models to entry-level vehicles under 100k RMB.

By bringing the World Model solution, previously exclusive to luxury cars, to new vehicles starting at 86.8k RMB, Leapmotor has broken the cost barrier for advanced driving assistance, driving the widespread adoption of 'point-to-point' functionality.

* 2. World Models outperform traditional end-to-end solutions in understanding physical laws.

Unlike end-to-end behavioral cloning (a 'black box'), World Models possess cognition of 3D space, physical laws, and causal relationships. They can perform 'mental sandbox' simulations of parallel worlds, making decisions more predictable and safer.

* 3. Extreme distillation technology solves the reliance of high-end AI models on massive computing power.

Through knowledge distillation and neural network architecture optimization, Leapmotor successfully deployed a World Model on the Qualcomm 8650 platform with only 100 TOPS of dense computing power, achieving an engineering leap by running advanced algorithms on low-cost hardware.

* 4. Leapmotor adopts a multi-path intelligent strategy to adapt to different product tiers.

Applying one-stage end-to-end, VLA (Vision-Language-Action), and World Models to different vehicle models reflects their fundamental exploration of returning to the essence of the car as a mobile robot—interacting with the physical world.

Key Quotes

* The greatest value of a World Model is that it allows for 'what-if' inferential interaction, understanding the causal chain between actions and consequences, rather than merely discovering statistical correlations. * Rules are like 'poison'. They solve problems quickly, but for every rule you write, you'll eventually need ten thousand more, resulting in a 'spaghetti code' algorithm that leads to inconsistent driving styles and a jerky user experience. * Extreme distillation, deploying a World Model on a platform with 100 TOPS of (dense) computing power. This is something even Musk hasn't achieved. * A good AI driver doesn't need to be deliberately tested.

AI Score

88

Website qbitai.com

Published At Today

Length 3650 words (about 15 min)

Tags

Leapmotor

World Model

Autonomous Driving

Model Distillation

End-to-End

Related Articles

* Major Update! Jensen Huang's First Keynote of the Year: The World is Undergoing Two Platform Revolutions! New Rubin Computing Architecture Debuts! First Autonomous Driving AI Open-Sourced, AI-Native Application Framework Released! Grok 5 Parameter Scale Revealed * Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * AI Starts to "Take Action", Alibaba's Qwen Leads the World * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * Google's New Paper Crashes Memory Stocks! 6x KV Cache Compression, Netizens: Silicon Valley Becomes Reality * Context Graphs: The Trillion-Dollar AI Frontier Trending on X – What They Are and Startup Opportunities * Four Steps of AI Transformation: Individual, Organization, Product, and Business (Part 2) * Zhipu GLM-5 Technical Deep Dive: Full Compatibility with Domestic Chips Like Huawei Sparks Global Discussion and asynchronous reinforcement learning infrastructure, while completing full-stack adaptation for domestic chips.") * CES 2026 Challenges: Exploring the Bubbles and Non-Consensus Behind 50 AI Projects * #426. Deep Dive with Google Legend Jeff Dean: The Evolution of Gemini, Hardware Co-design, and the Future of AI Programming HomeArticlesPodcastsVideosTweets

Leapmotor's New 86.8k RMB Car Democratizes Point-to-Point...

查看原文 → 發佈: 2026-03-28 14:16:49 收錄: 2026-03-28 16:00:33

🤖 問 AI

針對這篇文章提問,AI 會根據文章內容回答。按 Ctrl+Enter 送出。