How the Next Era of Industrial AI is Transforming Manufacturing, Dr. Ye Huang at Hannover Messe 2026

Ajouté : 2026-05-05

[00:00:00]that need to be taken. We heard about infrastructure to make industrial AI work and in general different components and partners need to come together but just a few companies are offering the full stack engine as we call it. Well, Alibaba does. Thus, they are introducing the 5A cloud framework to you today. And by they I actually mean the general manager of solutions and products at Alibaba Cloud Central Europe. He has lived in Europe for over 20 years. He told me he holds a PhD from Switzerland.

[00:00:34]And with this experience, he's leading the technology teams for Ali Cloud and Quen, focusing on the European market.

[00:00:41]So, please give a hand for Dr. Hi Huang.

[00:00:56]Good afternoon. Thanks a lot for the introduction here and good to see you guys here. Um my name is Y. Um it's pretty easy name to memorize is you never make someone wrong here. I'm leading the technore team for the Alibaba cloud and also the QN team our foundation model team in the central European part. So it's my great pleasure here to be here to presenting our vision here on one hand also our cutting edge progress our technologies.

[00:01:26]So before the starting of this session here I wanted to give you a quick uh recap where we're standing as moment here and from our perspective which is critical moment from the generative AI into the AI agents. So it's pretty much different compared with one year and a half ago where people are being already excited about the how AI can generate the content for us which is require a very powerful model at the time.

[00:01:54]Nowadays if you're looking as where we're standing here we're looking at the the whole technology stack which can cover to make the AI to move. So you're going to have the brand, you're going to have also the body who are very powerful here which require the uh the technical leaders to provide not only the very powerful financial model as well as the whole food step from carving from a applications into the platforms financial model of course themselves or the cloud infrastructure the platform the data uh infrastructure as well as the infrastructure which are dedicated optimized for the AI workload. cloud which is a standing point of Alibaba group nowadays and also where the dedication of the Alibaba cloud. So we hold the value proposition to providing the 5A cloud. If you're looking at the whole framework of the five cloud concept, we're dedication to make the AI innovations to bring up the cutting edge cloud infra the cutting edge foundation model allowing our customers to make the evolution by themsel is always being equipped with the cutting edge technologies never being fall behind.

[00:03:06]So we're dedicated providing any stack which with which definition where no matter the infrastructure being deployed from cloud into endpoint to the edge as well as a cloud not only dedicated for the infrastructure but also dedicated for the AI workload allowing a customers to utilize this kind of ending stack to bring to build up the workload by themselves freely which is crucial nowadays.

[00:03:32]We're committed to provide the any time which we have a team being distributed all around the world and to allow customers to be engaged with us anytime so we can bring up the efficiencies have always always on services and we also committed to bring the uh anywhere infrastructure which are deployed anywhere of the world including Asia including Europe including US and also North uh American and also South Americans to in to empower customer with the experience with seamless technology and also the services.

[00:04:08]So having said that I want to first of give you a bit glance about the portfolio of function model which is brilliant nowadays for the AI workload.

[00:04:18]So we have our language model the Korean model uh which many of you already been heard hopefully already which are equipped with the best performance which is like Q max. We also have a balanced version for the plus and also a very speedy f version for the flash part. The differentiation of the query model our language model here is from the day has been born or equipped it with multimodata capabilities from day one which means it's not being powerful on the language processing by themselves.

[00:04:50]It's also been very focused on the multimoda models. That's why we're providing the models to supporting the virtual and also with the ominino as well as with the image and the video part which are already being natively being supported from language model from day one. We also bring up our multimoda models which are capable processing the uh um video generation and video editing and image generation or image editing as well. We also have our for model which are very focused on the voice processing part understanding species.

[00:05:28]I want to pick up the two of the models because the the whole family is quite big. I only pick up the models which are being update in the recent weeks. The first one we want to point out here is the cool 3.6 models here. And the reason why I want to pick this one here because of three major features which are especially being adapted with the requirement from the enterprises. The first one is our unmatched long contact supporting features. We are capable of supporting from 3.6 Six models, one million of tokens. Which means if looking at enterprises, you can drop a documentation compressed of 1.5 words in one shot with a model to bring processed and which is pretty unique feature nowadays which have been being especially being fine-tuned for enterprise utilizations.

[00:06:21]The second feature of the 3.6 six model nowadays which was released roughly about two or three weeks ago is focused on the agenic coding here. So which means we're trying to make the best model on the mathematic part and the model himself is not just to be able to answer the question from software engineering part they are capable of to processing the tasks of the engineering by themselves independently. So it's not like behaving as the best consultant for the engineer but like best assistance for the engineering in the coding part.

[00:06:57]The last things I want to mention here is multimodella intelligence of the current 3.6 models which allows the cross modella reasoning which is the key nowadays for the enterprise utilizations. So apart from this very I got to say very complex uh leading SOTA benchmark you're more than welcome to check out by yourself and which are very confident on that part the something I want to highlight here is the after the release of the 3.6 of pre models here the open router which is very world leading of the AI aggregator nowadays has been showing already the daily utilization of queen model right after this one being released have reached 1.4 4 trillion tokens daily. So that's showing the performance of the model.

[00:07:46]But that's also shows how the community how enterprise nowadays they're committed using AI to supporting their daily work. The workload is dramatically high nowadays.

[00:07:58]The second model which I wanted to to to show with you here today is our cutting edge status of the uh omni model from Qin 3.5 which is have some special features also especially for the manufacturing customers and also for the automotive customers. So, so current 3.5 Omn models has been equipped with the latest supporting lady supporting for the language, latest supporting for the text, voice, audios and video part which allows a very long context 256k context which means you can drop 10 hours of audio in one shot in the mod or one hour or one hour video in one shot in module lens one to be processed. So we have been performed over 215 soda tests as moment and you can see has been showing a very leading positioning in the uh automotive driving uh fields reasoning and also dialects. So the omn channel has been very been suitable for the viber coding if have been test by yourselves which means you can drop a video here and the the model itself going to generate the front end and also the python coding for theelves to turn a video into a page for you straightforwarded by the automatic coding part. is also being very much fitting into the real-time interactions because it can give the interactive on the-fly feeling on voice control and emotion control on voice which perfect for the customer service scenario examples here. So you are more encouraged to check out these two models by yourself. And one things I also want to mention here is apart from language model here is K is not just focused on the enterprise supporting part and we are very proud to be the lean model on open source. So we are being ranked as number one as embracing the open open source communities and which means we have over 300s models has been open source across different features across different sizes which has been widely being adopted by the engineers which is of the core and foundation for us to step into the global market. We have already occupied if you look at the number from December of last year 1 billion of downloads which means 20% worldwide opensource downloads overall into the global market and we showed how the cream model has been loved by the engineers by the heart and nowadays we're also very proud to see as you can see the number is very impressive 200 thousands of models which are diverted from the kin models were based on kin architectures and which people love model and they made their own model on top of the query model themselves. So this all data you can check out from the hugging phase and also from the open leading dashboard we're showing the love from the community here. Apart from the dedications into the the language model, I also wanted to quickly to talk about the the video going to play out uh our multim models which called one which had been showing and the 2.7 being released from the 3rd of April. So our mind model are perfectly being utilized into the video generations as well as with video editing part. It had very native support from the text to video synchronizations which is crucial nowadays for the enterprise to use this ones to for the video generations which also we also allows very precise frame control by editing the videos. So you don't have to whole run the video generation time time again you're going to save the cost you're going to save the time for the enterprise when adding the video has been autogenerated by the model themselves.

[00:11:54]The vinyl mos also are perfect for the perfect for the image generations and also the video the image editing as well which is nowadays also very important for the department for example the digital marketing right which all enterprise nowadays kind of need it so as soon as the one model has been released I do believe on s of April we have been already being adopted by the pixart which is one of the world leading video AI creative platforms nowadays worldwide as a default model for the enterprise tactto video utilizations and also as a default model for the AIC social platforms as well showing the one model has been also being very popular nowadays by the uh by the AI creative platforms.

[00:12:42]Another things which I want to mention here which is not on the slides today because it's so new like I think one week and a half ago we just announced the happy horse the another very leading video uh models has been being ranked from the uh artificial analysis dashboard are also owned by the Alibaba cloud as well and also it's award from Alibaba group and five days ago we have launched another heavy horse family the heavy oyster is our first the B model here. So all this kind of kept innovative foundation models being rolling to the market has had been showing the AI innovations from the Alibaba group and also from the Alibaba cloud.

[00:13:24]So much about the foundation model here.

[00:13:26]I wanted to quickly talk about the also the platforms we're having here. So our mass as a service or model studio platforms are supporting not only in-house being made the function models of the QN family and also with the F family. We're also supporting the partner models Kimi uh for examples minimax which also popular being utilized by all enterprise customers.

[00:13:52]They're all have been supporting by the our model studio being deployed around the world. So nowadays model not only available from our APC regions like Hong Kong and Singapore it be also been deployed by end of March also in our friend regions and we can also deploy when the our model studio into different regions [snorts] uh uh soon as well to show the global uh deployment and also the long-term commitment for the market apart from the models apart from the mass I want to grab you a glance for a second to dive in bit deep into the info structure part it might be tech but it show the real innovation been coming from so if you're familiar with the cloud part you might be familiar with all the infrastructure uh the so-cal yas platform here but things I want to point here is we also made tons of AI dedicated innovations into the infrastructure on the IDC level here so we have a penu uh AI server which our in-housem made AI server comprised of 128 nodes inhouse being A in order to make the dedicated compute enhancement on the AI server part we have the internet based AI hyper network also inhouse being made which are together with our AI AI server which build up the linking AI clusters in English means which means strong and beauty horse. So all these kind of AI um dedications on the infrastructure part are the foundations for the um or the foundation model we're running here because when they're running together on the under laying of the of the infrastructure they can allow Alibaba cloud and Alibaba group to build up GPU clusters over 100,000 GPUs where still we keep a very high performance um and also a still very high of the time for training. So the number kept over 99% even when the GPU clusters go beyond 100,000 cars which is a very crucial not of mathematic problems but also an engineer challenge nowadays. So this will allow eventually us the Alibaba group and Alibaba cloud can can bring up the our model uh inference efficiencies to up to 5 4.5 times and also to bring up the service capacities also significantly which will make first of all the performance of infrastructure to be very competitive and then that will also make the cost per token also very controllable which is nowadays very critical. momentum for the enterprise to adopt the AI technology in their daily work.

[00:16:36]So much about the model and the platform itself and also the infrastructure. I also want to mention here we also are dedicated to bring up the a anytime services. So nowadays we have the offices across world including local office service teams. If you're looking at here for the European region for example we have the Frankfurt office, Munich, Armstand, Zurich, Paris, London.

[00:17:00]So the teams of the Alibaba cloud teams has been distributed across Europe showing the long-term commitment to run together with customers providing the office on services with customers not only team and also services we are also being committed to providing the infrastructure also across world. So nowadays we're operating 29 regions in a global scales and uh total 94 availability zones which means large data centers at this moment. So for example in Europe we have three uh data centers in Frankfurt but in coming we have the fourth ones coming up as soon and the capacity is so large which equal to the sum of purees together and in this month we're going to launch our par regions which also compress of three availability zones also in Paris and later this year we're going to launch our also the lesson regions which also been comprised of a couple of the data centers also in the lessons so which will build up the infrastructure foundations for the European mainland which also will be in in addition with our regions also in UK as well show the the competent we're having here so somewhat talking about the technologies and also the service teams and infrastructure I want to also share you about how we're looking at how the AI can shaping or can reinvent our daily utilizations so we do believe the AI can reinvent first of all the hardware or let's the physical world of products. Secondly, we do has been has been seen have been witnessed the AI can also can reinvent the user experience and third we also have been seen the AI can also reinvent the business efficiencies.

[00:18:46]So let's be looking at some of the things we're being seen together. I think I don't have I've met the mid made advertising here right if you look at the old robots nowadays uh all over the Hanuma method here uh we can see the robots they can do the crazy dynamic movements right for dancing from kung fu which are being empowered by the AI part and also for the Alibaba group we are being made our also testing on the wearable devices for example the quinn classes nowadays being rolling out by the Alibaba which can supporting the real time human interactions can supporting people the experience in the in the navigations in the e-commerce in the payments in the daily life. So we have been witness and we've been testing water the temperature by ourself to see how AI can change the physical world can change the physical products.

[00:19:41]We have been also been cooperating with partners to see how the AI can change the experience. For example, the BMW group have launched the strategy partnership of the Alibaba group to build up the AI engines which are empowered by the Quinar to enhance the user experience driver experience for the Noi class for the BMW vehicles in the Chinese market. So specifically the BMW group have built up the AI agents uh called car genius and also travel companion in order to enhance this kind of experience for the vehicle control for examples and also the planning the very complex dialogic management with drive between driver and the vehicles on the road which can eventually after different testing with the huge amount of parameters and experience can bring up the the the final result. for example 99% of wake up occurrences as well as compute of the power boost for 20 times.

[00:20:42]So that's which we have been witnessed how AI can change can also improve the user experience.

[00:20:49]We have been also have the honor to cooperate the partners to change the business efficiencies by the AI for examples with one of leading platform in GBS platforms the hypert based companies who also being recognized by the Gartner Metric Quron here. So they are dedicating providing their platforms for the men on the loop for providing this kind of AI platforms to change the uh the all the workflows for example procurement to pay and order to cash. So the hapt tools has been adopted the kin model by themselves in order to enhance this kind of efficiencies. For examples, the 90% of the men on loop uh can be can be facilitated by the AI capability from the Q models 90% of the of the process can go straight forwardly. So when the habit go engage with customer they can say the after the investment for the six months to nine months all demands will be will be compensated by the real results and they also have be calculated by the result eventually like per person intelligent documentation processing per day has been significantly improved from 50 to over thousands showing how business efficiency has been improved through the time as well. So here because of the time limitation I can only share with some of cases but we have been have been honored to cooperate with the enterprise across world uh trust by the enterprise and we also have the whole booth here in the hall 14 um booth 46 if you have time you're more than welcome to drop by to have a conversation with me and who's my team here as well. So having said that so I would like to appreciate a lot for your time your dedication listening to the conversations and we're very much looking forward for the long run together and for Australia partnership will for you here. Thanks a lot for your time. [applause] >> Thank you very much y Thank you. Yeah you can give it to me. I was uh thinking about what Joan K Rowling maybe have in mind nowadays that everything is in place and how the Harry Potter films would look like if they would use your models and technology.

[00:23:03]>> Yeah.

[00:23:04]>> And how maybe the writers of James Bond are envisioning the future now.

[00:23:09]>> We're building future together with partners eventually.

[00:23:12]>> We'll see. Thank you very much. We find you in hall 15 booth 46. Thank you very much. Yay.

[00:23:17]>> Thank you. Thank you.