灵波开源，中国 AI 的另一种出海叙事
2026 年 9 月 4 日，又一年柏林国际消费电子展 IFA 大会开幕，会展中心的创新舞台座无虚席。
IFA 2026 的开幕主题演讲题目是《个人 AI 时代：想象力的未来》，主讲人是 AMD 高级副总裁、计算与图形事业部总经理 Jack Huynh。
作为消费电子圈每年最重要的场合之一，台下所有人都在等 AMD 拿出年度最重磅的内容。Huynh 的前半段演讲照例讲硬件：新一代 Ryzen AI Max PRO 处理器，AMD 历经多年打磨、专为个人 AI 设计的笔记本「Sundance」。
到演示环节，他话锋一转：「如果说，用一句 prompt 就创造一整个世界呢？」
他双手托着笔记本，屏幕上没有播放键，镜头平稳地向前推进，穿过中国小镇的街巷，又走上欧洲乡村的小路，光影随着视角移动自然变化，世界一路向前铺展，仿佛没有尽头。
这不是一段提前渲染好的视频，而是一个正在实时生成、可以交互的三维世界。
过去一段时间，这类能力更多出现在实验室和论文演示里，能以这么高的完成度，出现在消费电子展的主舞台上，还是行业第一次。
演示结束时，Huynh 对着话筒说了句：「This is the future of Personal AI。」
很多人不知道的是，驱动这个演示的世界模型 LingBot-World 2.0，正来自蚂蚁集团旗下的具身智能团队灵波科技（Robbyant）。
三个月后回看，那场演示只是一个开始。从高校实验室到芯片厂商的开发套件，从独立开发者的工具到垂直行业的工作流，LingBot（灵波）系列模型正在进入全球具身智能的技术底层，成为很多团队创新的起点，甚至还进入了红海珊瑚礁的研究现场。
01
成为学术研究的参考基准
能被 AMD 选上舞台，不是运气。在此之前，灵波已经在全球具身智能的学术圈成了一把隐形标尺。
世界模型赛道有个公认的难题：物体恒存。简单说就是镜头离开一个物体，再转回来，生成模型还能不能把它还原出来？
南洋理工大学、清华大学和普林斯顿大学的联合团队，把这个难题当成测试世界模型记忆能力的试金石。世界模型需要利用此前的画面，维持物体和场景的连续性；生成时间越长，保存历史信息的缓存也可能越大。研究者想做的，是压缩缓存的同时，保留后续生成真正需要的信息。
团队比较了 LingBot-World v2、LongLive、Self-Forcing 和 Causal Forcing。在特定的对象找回与场景重访测试中，LingBot-World v2 利用完整历史信息生成的画面，与目标画面的 DINO 特征相似度约为 0.68，其他三个候选约为 0.26。
这个结果直接决定了他们后续的研究方向：放弃其他候选，直接在灵波的底座上做记忆压缩优化。最终提出的 DeCoPrune 方案，把历史缓存 token 砍掉了 85%，生成速度翻了四倍多的情况下，画面相似度还能保住 0.67。
类似的故事，在具身智能论文里反复出现。
斯坦福大学视觉与学习实验室 10 月发布的 OpenWAM 项目，初衷是给世界 - 动作模型（World-Action Model）做一套统一的对比框架。过去各家模型各写各的代码，架构不兼容、接口不统一，想比个优劣得先重写半套环境。他们找了一圈现成的实现，最后把 LingBot-VA 的 Transformer 架构、文本编码器和分词器整套拿了过来，作为整个框架的参考实现。
韩国 KAIST 与成均馆大学团队的联合研究成果 FastOPD，则用 LingBot-VLA 作为「教师模型」，为约 4.51 亿参数的学生模型提供监督，最终让学生模型单步推理的平均成功率由 35.3%提高到 51.2%，增加 15.9 个百分点。
这些研究延续了灵波已经完成的不同部分。有人使用模型能力，有人沿用工程实现，再将自己的训练资源和研究方法投入其中……而当一个领域的新研究都用灵波来做实验、当别人要证明自己的方法有效首先得跑过灵波的基线时，灵波也不再只是一个参赛者，更是这个赛道的度量衡。
02
开源社区：全球开发者的二次创造
能被学术圈使用，依然还是小圈子里极具门槛的事，全球开发者的二次创造，为开源产品带来了第二次生命。
LingBot-World 2.0 除了 14B 的完整版，还开源了一个面向普通开发者的 1.3B 快速版，面向普通开发者。很快就有人拿去试手。
爱沙尼亚的独立开发者 Kaarel Kaarelson，就是最早把灵波的性能潜力榨干给所有人看的人之一。他在生成式 AI 圈里小有名气，擅长给开源模型做推理优化，别人跑不动的模型，到他手里总能快上一截。
9 月，他拿到 LingBot-World 2.0 的 1.3B 版本代码后，一头扎了进去。他先在云平台租了一张 RTX 5090，32GB 显存，跑官方的 1.3B 版本。在他的设置下，原始实现生成一秒钟、16 帧的视频需要 2.68 秒。用户持续移动时，画面生成赶不上播放。
于是，接下来整整两周，他从把 VAE 精度从 FP32 降到 FP16 开始，替换注意力算子为 SageAttention，线性层做 FP8 行式量化，用 torch.compile 把十几张计算图合并成一张，再把层间的 CPU-GPU 同步从 110 次砍到 2 次，最后做算子融合。
每改一步，他就记一组耗时，连每个算子离硬件理论峰值差多少都算得清清楚楚。
最终结果出来的时候，连他自己都意外：同一张显卡，原版 6 帧每秒，他优化完跑到了 16.1 帧，速度翻了 2.7 倍，画质几乎无损。
他把对比视频和代码一起放到 GitHub 和 X 上。左边是原版卡顿的龙穴场景，右边是丝滑流畅的 16 帧版本。那条帖子很快在圈子里传了起来。科技媒体 Prismix 专门写了报道，认为它在动态模拟上已经超过了闭源的 Genie 3。
Mac 用户群体也没有缺席。PandacatAI 完成了更底层的 MPS 后端移植，用 M4 Pro、48GB 统一内存，配上 8 位量化的文本编码器和磁盘提示词缓存，让 Mac 用户也能用 LingBot-World 本地生成世界。
开发者 Don Allen III 则为 LingBot-Map 做了一个完整的 Mac 桌面封装：用户把视频拖进窗口，就能在本地生成三维点云。全程可视化进度，最终导出点云文件。
开发者的热情攒到一定程度，产业端的官方主动适配就成了水到渠成的事。
03
产业生态：宿敌之间的默契
要了解合作伙伴们究竟如何看待灵波，英伟达与 AMD 这对宿敌的默契，或许是最好的例证。
熟悉 AI 硬件圈的人都知道，两家的生态向来是独立发展的，尤其是世界模型，两家都会优先推自家投资或者合作的模型：英伟达有自研的 Cosmos 系列；AMD 则在投资了 World Labs 后，双方官宣开展技术和 go-to-market 合作，苏姿丰与李飞飞更是在多个公开场合互相背书。
灵波是少有的例外。
NVIDIA 在其开源高性能推理库 FlashDreams 中，将 LingBot-World 与自家的 OmniDreams 并列，作为三大官方集成模型之一，专门提供相机控制的图像转视频服务。在 SANA-WM 的研究论文中，英伟达更是直接将 LingBot-World 称为「大规模工业级基线」，用它来衡量自家新模型的质量水平。
另一边，AMD 不仅在 IFA 2026 这样级别的消费电子盛会上用 LingBot-World 2.0 演示「个人 AI 的未来」，其 ROCm 软件生态也在跟进 Micro-World 系列的优化，并公开评测中认可 LingBot-World-V2-1.3B 的 WBench 质量分达到 81.22，位居同类第一。
芯片厂商之外，工具链的跟进同样迅速。
推理框架层面，SGLang 宣布对 LingBot-World 2.0 提供 Day 0 Support，成为官方首发的推理加速方案；vLLM-Omni 随后也相继合入了视频生成与实时世界会话的支持。Hugging Face 旗下的机器人开发平台 LeRobot，更是将 LingBot-VA 纳入 v0.6.0 核心版本，作为官方推荐的世界动作模型策略之一。
视觉开发者熟悉的 timm 模型库，也在 2026 年 8 月正式收录了 LingBot-Vision。从此，全球的计算机视觉工程师只需一行代码，就能加载灵波的视觉骨干网络来做迁移学习。
云厂商也没落下。AWS Labs 的开源世界模型部署项目，直接把 LingBot-World-Fast 设为默认的 EC2 部署路径，并整理好了从环境到服务的全套流程。很多开发者第一次在云上部署世界模型，第一个接触到的选项可能就是灵波。
从云端部署到端侧芯片，从训练框架到推理 runtime，一条围绕 LingBot 的完整产业工具链正在自发形成。
04
从机器人的眼睛，到红海的海底
如果说前面所有的故事，都还在具身智能的射程之内。下面这一个则不同。
没下过水的人想象珊瑚礁，大概类似于一张明信片：清澈的蓝色海水，成群的热带鱼，潜水员游动在光斑里。
但在瑞士洛桑联邦理工学院（EPFL）环境计算实验室的人眼里，珊瑚礁是另一种画面：一段手持相机拍下的、随水波晃动的视频，以及怎样把它变成一张能数清珊瑚数量的三维地图。
这个叫 DeepReefMap 的项目，研究的是红海北部亚喀巴湾的珊瑚。作为全球最特殊的一批珊瑚，它们对海水升温有超乎寻常的耐受力，被视作气候变化下珊瑚存续的希望。
为此，EPFL 跨国红海中心连年在以色列、约旦、吉布提沿岸做考察，通过让潜水员带着 GoPro 潜水拍摄视频，试图完成成整片区域的三维重建。
但水下的麻烦比陆地上多得多。水波晃出的焦散光斑、悬浮颗粒造成的后向散射，都会干扰特征匹配。一些重建方式需要同时读取整段视频，帧数增加，显存需求与团队用于后期处理的也会同步增长。
转折点发生在 2026 年 9 月 24 日。
那天，DeepReefMap 发布 1.2.0 版本，增加了 LingBot-Map 等重建后端。在官方描述中， lingbot_map 权重公开，Apache-2.0 许可，做逐帧流式处理时，显存始终平稳，是「处理长视频或使用显存较小的 GPU 时的最佳选择。」
原本为机器人导航设计的空间理解能力，就这样走进了海洋生物学实验室。
一定程度上，这就是开源最迷人的地方：当一个模型足够开放、工程底座足够扎实的时候，使用者自会自己替它找到千万种用法。
05
尾声：中国 AI 的另一种出海
很长一段时间里，中国 AI 的全球化叙事总绕不开应用出海——把国内跑通的产品模式复制到海外。
但灵波走了另一条的路：技术底座出海。
依靠 GitHub 仓库里的 Star 与 Fork、全球开发者自发的适配与移植、每一篇论文里的基线数据、每一次性能优化里的实测结果……这些分散的、自发的选择累积起来，最终构成了模型真正的生态影响力。
而这一次「 出海 」最深的抵达，是字面意义上的一片海——红海北部的亚喀巴湾，一段晃动的水下视频，和一群在升温的海水里依然存活的珊瑚。
一个开源的模型，就这样成为无数创新发生的土壤。
*头图来源：DeepReefMap GitHub 主页
本文为极客公园原创文章，转载请联系极客君微信 geekparkGO
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
It's another one from China's AI.
On 4 September 2026, another year of the International Consumer Electronics Fair in Berlin opened the IFA General Assembly, where there was no gap in the innovation stage of the Centre.
IFA 2026 ' s opening keynote address is " The Personal AI Age: The Future of Imagination ", with the lead speaker being AMD Senior Vice-President and General Manager of the Ministry of Computing and Graphics, Jack Huynh.
As one of the most important events of the year in the consumer electronic ring, everyone is waiting for the AMD to come up with the most heavy weight of the year. The first half of Huynh's speech is normally about hardware: a new generation of Ryzen AI Max PRO Processors, AMD's notebook "Sundance" designed for personal AI after years of grinding.
And at the demonstration, he said, "What if, by a single sentence, the whole world is created?"
With his hands on his notebook, there was no play key on the screen, and shot was moving forward in a smooth way, through the streets of small Chinese towns and into the streets of the European countryside, where the light moved with his eyes, and the world moved forward as if there was no end.
This is not an early replica, but a real-time, interactive, three-dimensional world.
Over time, such capabilities have been more present in laboratory and paper presentations, and at such a high degree of completion, in the main stage of consumer electronic exhibitions, and for the first time in the industry.
At the end of the presentation, Huynh said, "This is the truth of personal AI."
What many do not know is that the world model that drives this demonstration LingBot-World 2.0 is coming from a smart team under the ants group's banner, Robbyant.
Three months later, that demonstration was just the beginning. The development packages from tertiary labs to chip manufacturers, from the tools of independent developers to the vertical workflow, LingBot (LingBot) series models are entering the global, smart technological base, becoming the starting point for many team innovations and even entering research sites on the Red Sea coral reefs.
01
As a reference for academic research
Can be picked by AMD on stage, not luck. Prior to that, Lingbo had become an invisible ruler in a globally intelligent academic world.
There's a recognized problem with the world model track: the survival of objects. Simply put, shot leaves an object and then turns back. Can the model be restored?
The combined team of the South China University of Technology, the University of Qinghua and the University of Princeton used this problem as a test case for the world ' s model memory. World models need to use previous images to maintain the continuity of objects and scenes; the longer they are generated, the larger the cache is likely to preserve historical information. What researchers want to do is to compress cache while retaining follow-up to generate really needed information.
The team compared LingBot-World v2, LongLive, Self-Forcing and Capital Forcing. LingBot-World v2 uses images generated from complete historical information to retrace the DNO features of the target image by approximately 0.68 and the other three candidates by about 0.26.
This results directly determine the course of their follow-up research: abandoning other candidates and optimizing memory compression directly on the base of the Leap. The final DeCoPrune project, cutting down 85% of the historical cache token, has more than quadrupled the rate of generation, and the image similarity can be maintained at 0.67.
Similar stories, repeated in smart papers.
The OpenWAM project at the Stanford University Visual and Learning Laboratory, October publish, was originally designed to provide a unified comparison framework for the World-Action Model. In the past, various models had developed different codes, with incompatible structures and inconsistent interfaces, and half of the environment had to be rewritten before it was better or worse. They looked for an off-the-shelf set, and finally brought over the LingBot-VA Transformer architecture, the text encoder and the syllable set as a reference for the entire frame.
The results of the joint KAIST research with the Cheng capital university team in Korea, FastOPD, used LingBot-VLA as a "teacher model" to provide oversight of student models with approximately 451 million parameters, ultimately raising the average one-step success rate of the student model from 35.3 per cent to 51.2 per cent, an increase of 15.9 percentage points.
These studies are a continuation of the different parts of Lingbo that have been completed. Some people use model capabilities, others follow engineering and invest their own training resources and research methods... And when new studies in one field are experimented with Rimbaud, and when others have to run through the baseline of Rambo in the first place to prove that their methods are effective, it is not just a contestant, but a measure of the track.
02
Open-source communities: global developers ' secondary creations
The ability to be used by the academic community is still a very high threshold in small circles, with a second creation by global developers, bringing a second life to open-source products.
LingBot-World 2.0 In addition to the full version of 14B, an open-source 1.3B fast-track version for ordinary developers and for ordinary developers. Soon someone will try it.
The Estonian independent developer, Kaarel Kaarelson, was one of the first to drain the performance potential of Lingbo to everyone. He's a little famous in the generation of AI, and he's good at doing reasoning optimization for open-source models, models where people can't run, and he can always get to him.
In September, after he got the 1.3B version of LingBot-World 2.0, he stuck his head in it. He first rented an RTX 5090, 32GB VRAM from the cloud platform, running the official 1.3B version. Under his settings, the original realization produces a second, 16 frame video in 2.68 seconds. When the user continues to move, the image generation does not match the play.
So, for the next two weeks, he started from moving the VAE precision from FP32 to FP16, replacing the focus algorithm with SageAttention, making a FP8 line quantization in linear layers, combining a dozen calculations with torch.compile, and synchronizing the CPU-GPU layer from 110 to two times, and finally integrating the count.
For each change, he counts a set of time-consuming times, even if the difference between the theoretical height of each algorithm and the hardware is clear.
When the final result came out, even he was surprised: the same GPU, the original 6 frame per second, he optimized to 16.1 frame, double the speed of 2.7, with almost no damage to the paint.
He places the contrast video with the code on GitHub and X. On the left is the original Carton's cave scene, and on the right is the smooth 16 frame. The post came up in the circle very quickly. Technical media, Prismix, wrote about it as having surpassed closed-source Genie 3 in dynamic simulations.
The Mac user group was also not absent. Pandacat AI has completed a lower level MPS backend port, using M4 Pro, 48GB unified memory, with 8-bit quantization text encoders and disk promptZ96Z, allowing Mac users to generate the world locally using LingBot-World.
The developer Don Allen III made a full Mac desktop envelope for LingBot-Map: by dragging the video into the window, the user can generate a three-dimensional cloud locally. Full visualization of progress and final export of cloud files.
The enthusiasm of the developers has gained a certain degree, and the official initiative of the industry has become a matter of water to the canal.
03
Industrial ecology: convergence between rivals
It's probably the best example of how partners look at Lingbo, In Weidar and AMD as a match for rivals.
As those familiar with AI hardware circles know, the two ecosystems have traditionally been developed independently, especially the world model, and both have given priority to their own models of investment or cooperation: British Weaverda has a self-research Cosmos series; and AMD, after investing in World Labs, has announced technical and go-to-market cooperation, and Sozefo and Li Fei have endorsed each other in various public places.
Lingbo is a rare exception.
NVIDIA uses LingBot-World alongside its own OmniDreams in its open source high performance reasoning library FlashDreams, as one of the three official integration models, which provide camera-controlled image-to-video services. In the SANA-WM research paper, Young Weidar refers directly to LingBot-World as the “large-scale industrial-grade baseline” against which to measure the quality of his own new model.
On the other hand, not only did the AMD demonstrate the "Future of Personal AI" with LingBot-World 2.0 at a consumer e-fest at a level like IFA 2026, its ROCm software ecology is also following up on the Micro-World series, and in a public evaluation, the WBench mass of LingBot-World-V2-1.3B was recognized as 81.22 and ranked first.
Beyond the chip manufacturer, the tool chain follows up as quickly.
At the framework level of reasoning, SGLang announced that it offered Day 0 Support as an officially launched reasoning acceleration programme for LingBot-World 2.0; v LLM-Omni was then successively supported by video generation and real-time world sessions. The Hubging Face robotics development platform LeRobot, which incorporates LingBot-VA into the v0.6.0 core version as one of the officially recommended world action model strategies.
LingBot-Vision was also officially recorded in August 2026 by a timm model library familiar to visual developers. Since then, computer visual engineers around the world will need only one line of code to carry the Lingbot's visual backbone network for migration learning.
The cloud factory didn't fall. AWS Labs Open-source World Model Deployment Project, which directly sets LingBot-World-Fast as the default EC2 deployment path, and has organized the full set of processes from environment to service. For the first time, many developers have deployed a world model on a cloud, and the first option to be exposed may be Le Bo.
From clouds to end-side chips, from training framework to reasoning runtime, a complete industrial tool chain around LingBot is being spontaneously formed.
04
From robotic eyes to the bottom of the Red Sea.
If all the stories ahead are in the smart range. The next one is different.
People who have not been in water imagine coral reefs, like a postcard: clear blue seawater, swarms of tropical fish, divers moving in the light.
But in the eyes of people at the Environmental Calculator Laboratory of the Federal Polytechnic Institute (EPFL) in Lausanne, Switzerland, coral reefs are another image: a video of a hand-held camera that moves with water waves and how to turn it into a three-dimensional map that counts the number of corals.
This project, DeepReefMap, is studying corals in the northern Aqaba Bay of the Red Sea. As the world ' s most special group of corals, they have an extraordinary resilience to sea-water warming and are seen as a hope for coral survival under climate change.
To this end, the EPFL Multinational Red Sea Centre has been conducting a series of missions along the coasts of Israel, Jordan and Djibouti in an attempt to complete the three-dimensional reconstruction of the entire region by allowing divers to take GoPro dive and film.
But there is much more trouble under water than on land. The back-scattering caused by the flashing of the water and the flashing of suspended particles interferes with the matching of characteristics. Some of the reconstruction methods require simultaneous reading of the whole video, increased frame numbers and a simultaneous increase in VRAM demand and team use for post-production processing.
The turning point occurred on September 24, 2026.
That day, DeepReefMap publish version 1.2.0, added LingBot-Map etc. to the reconstruction backend. In the official description, the lingbot map weights are made public, the Apache-2.0 permission, and VRAM remains smooth in the frame-by-frame processing, "the best option for processing long videos or using smaller GPUs of VRAM."
The space-based understanding capabilities that were originally designed for robotic navigation thus walked into marine biology laboratories.
To some extent, this is the most fascinating part of open source: When a model is sufficiently open and the project base is solid, the user will find itself a million uses for it.
05
End: China AI's other way out
For a long time, China's AI narrative of globalization has been insulated from applying to the sea — replicating the domestic run-off product model overseas.
But Lingbo went another way: the technology base went out.
Star and Fork in GitHub warehouse, the self-adaptation and transplantation of global developers, baseline data in each paper, and the results of each performance optimization... These scattered, spontaneous choices accumulate and ultimately form the true ecological impact of the model.
And this one of the deepest "off to sea" arrivals was a literally sea -- the Gulf of Aqaba, north of the Red Sea, a shaking underwater video, and a group of corals still alive in warmer water.
An open-source model has become the soil for countless innovations.
*headchart source: DeepReefMap GitHub homepage
This article is an original article from Polar Park.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.