Discussion keeps the world turning.
A year ago, open source AI was just tech jargon for serious coders.
Now, though, anyone can download, tweak and share these models.
So who benefits in that situation?
Well, some see innovation, others see risk.
But this affects the apps that you and I use and the choices that you and I will have.
One year after DeepSeek changed the game, where do we stand now?
Coming to you live from our studios in Beijing, this is Roundtable.
I'm Steve Hatherly.
Thank you so much for being with us today.
And for the show, I'm with Fei-Fei and Yuxia.
First up.
A year ago or so, open source AI sounded like something that you might hear in a very serious meeting between people who own too many keyboards.
But something has changed.
AI models are no longer just products sitting behind the walls of a few giant tech companies.
More and more of them are being opened up, downloaded, modified, shared, and built upon by people all around the world.
But if everyone can get their hands on the technology, well, then who comes out on top?
Does open source make AI cheaper or faster or more accessible?
And I wonder this too.
Can companies really make money by giving their technology away?
Or is it truly being given away?
And should some AI models perhaps not be opened up at all?
That needs to be discussed.
Maybe more importantly, what does it have to do with you or me?
So let's talk about it.
One year after DeepSeek put open source AI firmly in the spotlight,
Where are we now and what has changed over the last 12 months?
Feifei and Yusheng, good afternoon to you both.
Yusheng, I remember you and I were on the show last year when we were talking about open source, when it was just kind of a new term.
Mm-hmm.
Of course, open source itself is not that new, but it is kind of the idea that probably last year was the first time that get into our eyes.
And last year, you might have downloaded an open model to run a chatbot or test this feature.
Reasonability, that's what a lot of these coders would do and probably software engineers.
Today, the same kind of model can sit inside a company's AI system.
It can search internal documents for employees, help programmers write and debug their code,
Power customer service agents or analyze data from a factory
So developers can also take an exciting or I mean existing model and fine-tune it for a specific industry and feature and build an AI product around it instead of just training a model from scratch.
In other words, the change is that open source AI is moving up the technology stack from a model you experiment with to a tool you can actually build a business or service around.
Yeah, I think in the past when deep seek actually for a not a very tech savvy person like me, open source is kind of the term that could slipped very quickly because there are a lot of terms surrounding that model that they publish.
But what's really catching our attention now, you know, one year after DeepSeq really put open source AI in the spotlight is that we are seeing some report, in particular the Hugging Faces Spring 2026 report,
Which is the world's largest online community and platform for sharing and developing open source AI models,
They said China developed open source models accounted for more than 40% of all model downloads worldwide.
That is even bigger than the United States.
And another data coming from the Chinese Ministry of Industry and Information Technology also said that accumulated downloads of Chinese AI models have surpassed 10 billion.
Yeah, there's lots of reporting going on.
There was another report on a platform that allows users to access different AI models through one interface.
It's called Open Router.
They were talking about the fact that the top five most
Called models are now from Chinese companies.
Xiaomi is one of those.
DeepSeek is one.
Tencent is one.
The Xiaomi one, it's called Mimo, I think, the v2.5 version 2.5.
That one, for example, reportedly reached 10.5 trillion tokens in weekly calls.
So it is a sign that open source is becoming more common and that China is leading the way, apparently, as well.
Understand this called as most used in for now we'll explain it later but the thing is that it is just widely used worldwide not only in china of course the the user base here in china is huge
But the thing is that in these kind of international and global platforms is also showing a big trend that these Chinese AI models are being used by like engineers and also developers worldwide.
And that brings me a new question, really, for me, actually.
You know, open source.
I'm assuming if there is open source, there is closed source.
So what does open source mean?
I mean, does it mean that it's open to everyone like you and me that is not really a developer or business?
Yeah, let me like use a kind of analogy so that you can maybe understand easily.
A closed source AI model is like going to a restaurant, you order a dish, the chef prepares it for you and you eat it.
You don't get the recipe or the ingredients, you simply just use the finished product and you pay for them.
An open source model is more like being given the recipe and also ingredients.
The restaurant offers you everything.
You can take them home.
You can cook the dish yourself and change the recipe and build something new from it.
It's a meal kit.
Yeah.
Kind of.
So that's something that I can work on if I have the know-how myself.
Yes.
So in the AI world, this generally means developers make important parts of the model available so that
Other people and developers can download it and run it and modify and build upon it, depending on the specific license, like what kind of other feature they are allowed to use.
And this is where some terminology can get confusing because we mentioned calling.
Yeah.
Let's talk about the terms here.
Calling a model versus, I guess the opposite would be downloading it, right?
If I don't have a powerful computer capable of running a large AI model, I can simply just call it.
Which means that my computer sends a request to a model running somewhere else, maybe the server of this platform that you're using.
And then through an API or other interface, I use the model remotely and then pay according to how much text the model processes.
And then downloading...
So it's on a server somewhere.
So I can sort of analogy like making a phone call.
You can use it in the API.
Think of it like the waiter in the restaurant.
The waiter takes your order and then goes to the kitchen and tells the chef, the model, what to make.
And then once the model, the chef, makes it, the API, the waiter brings it back to you.
Your computer to a server somewhere that's doing the work for you and it gets delivered back to you.
But downloading is different.
Yeah, because a company such as a bank or energy company or a government agency may want to put a model on its own servers.
And that gives the organization more control over data customization and how the model is integrated into their own systems.
So you can think of downloading as like calling is like ordering the food and downloading is like taking the recipe and cooking it in your own kitchen.
Like catering on some level, right?
You're doing it by yourself.
Parameters is another term that's kind of a big deal when we talk about AI.
Yeah.
How powerful it is.
How big it is, how powerful it is.
It's interesting because some companies...
It was on July 27th, right?
That's when Moonshot AI...
Officially released the full model weights, we're going to have to talk about weights too, and tech support for Kimi K3.
That's the LLM from Moonshot AI, right?
Right.
It was in the news everywhere.
Really kind of huge LLM.
This the size of this, the parameters of this were bigger than anything I've ever heard of before.
Exactly.
And it is one of the kind that it is taking the lead globally.
So it says with 2.8 trillion parameters, Kimi K3 is an enormous model.
Its total parameter count is more than four times that 671 billion parameters in DeepSeq V3.
It is the, I think,
A consumer size model that DeepSeek is releasing from last year.
And this model is also one of the open models that helped
Put China's AI industry on the global map.
Parameters are kind of confusing because I don't really understand what it means beyond, and I don't need to, beyond knowing that it's kind of like the size of the model, the scale of the model, and somehow...
What it's capable of.
But just because we know that Kimi K3 has 2.8 trillion parameters and DeepSeek V3 only 671 billion, that doesn't mean Kimi K3 is four times stronger than DeepSeek V3, right?
And that's where it gets a little confusing, I think.
It's, like, another concept of how, like, when you're taxing to these Kimi K3, it's not using all of the 2.8 trillion...
Parameters is probably only using their activated 1 billion kind of thing
And that's particularly striking comparison because deep seek v3 itself was already considered a huge model when it was released and then this Kimi
Uses a mixture of experts architecture, meaning it doesn't activate all of their 2.8 trillion parameters for every question.
Yeah.
The more parameters, the more capable it is of completing complex tasks, right?
But for people like us who are just saying,
Ask him questions.
What's a parameter?
Yeah.
It's not going to need $2.8 trillion to complete that.
That's why smaller models are okay for us for everyday use.
And then here comes another like idea is that for other kind of international companies like chai jbt and cloud from anthropic we cannot make the same you know parameter count comparison because their company do not publicly disclose the size of their latest frontier models
And that is also kind of the idea explained
What open source is
Open source means we know how how much like parameter or how big it is and then uh closed sourced we don't know how it is
Yeah and i was
I was looking that up because i wondered you know why would
And and kimmy k3 that's open source right
Yeah yeah so why would
Why did this company make their model open source?
Why did DeepSeek do it last year?
But why do Anthropic and ChatGPT, for example, I think the earlier versions of ChatGPT, they made their parameter counts public, but I think it's kept private now.
So why do they do that?
And there's a couple of reasons I found.
One is security.
I think we're going to touch on that a little bit later.
But another is just competitive advantage because what they say is that if you release your parameter count, that's kind of a clue or it can be a clue to how your model was constructed.
Mm-hmm.
So companies keep that, that's one of the reasons they keep it private.
So for a company that's deciding that they're launching an open source model like Kimi, like Moonshot or DeepSeek, they're sort of, quote unquote, revealing partially, at least partially of their, quote unquote, secret.
Mm-hmm.
And that is also why I think for a long time when I saw it on the news, this company released their own AI models using DeepSeek.
So they're sort of basically using what DeepSeek released and then developing their own version of tools based on AI.
Yes.
But it is also important to mention that open source doesn't necessarily mean completely free.
The model ways are available, but using the model at scale still involves computing costs and also the license.
The license comes with commercial conditions, and we'll come back to what open actually means and how companies can actually make money from their open sourced models later.
Yeah, you're not simply giving it away.
There is a business model here.
Going back to DeepSeek though, what happened to open source after all that excitement?
Because that was kind of the big thing.
When DeepSeek came out, it dominated the tech headlines because people couldn't believe that they were able to construct a model like that.
At the cost that they did.
And it was also open source.
So that was the big news at the time.
But what happened in the aftermath?
One of the biggest changes is that open source AI appears to be moving from the kind of experimental stage from last year into something much more practical.
And the story is no longer simply about...
How these companies are competitive enough to build the competitive AI models, it is increasingly about whether those models are becoming part of the global AI ecosystem.
And there are some data is showing that the data from the US investment institutions, as many as 80% of American AI startups are now using Chinese open source models in their fundraising pitches.
And on average, more than 200 derivative
Models are being developed every day based on Chinese foundation models?
So thinking, for example, DeepSeq, that is basically my knowledge, it sort of has this very powerful and general purpose engine.
It is very smart and it can handle a lot of different questions.
But for instance, a small company or a city government that they don't have this generic engine, but it also needs one that knows its own local rules, knows its forms, knows its policies.
And that is when they take that engine from that deep sea structure and architecture and then adapt it.
They feed in their own documents, train it on local regulation, and tweak it so that it can give accurate answers about very specific and detailed things.
And that is what we are talking about when we're seeing these startups are using Chinese open source models.
Yeah, because I think a local government is a great example.
I think a bank is another great example.
A bank would, if a bank has a chat bot,
Well, the people who ask questions to that bank chat bot, those are going to be bank-related questions, right?
Exactly.
So you wouldn't just put DeepSeek as your chat bot.
You would take the open source DeepSeek or Kimmy K3, if you need it to be that many parameters, meaning if you need it to be that big and complex,
And you would tune it, fine tune it to your own needs so that it can answer questions specifically that customers of your bank would ask.
And this is about creating a better service for your customers and a better tool for your company.
Yeah.
At core, they all share the same powerful base, but each one is tailored to a unique or specific task.
And this is exactly what's happening with these platforms.
They take...
For example, the AI models from DeepSeek as the foundation and then reshape it to serve their own users and making the AI useful in real life, not in
Labs so where does the let's talk about the money now because I too when I hear open source I think it's being given away
It's free, and it makes me think, what a terrible business strategy.
But what a generous company to seek is.
Exactly.
How generous you are, but what a terrible business strategy.
But it doesn't mean that it's entirely free, and it also doesn't mean that there's no business model here.
As we just mentioned that it does not necessarily mean it is totally free and they will be a kind of ecosystem acquisition cost, which means that instead of thinking of the model itself as the final product, a company can see it as a way of getting millions of developers and businesses involved.
Into its ecosystem.
And once they're there, there are plenty of things that company can charge for.
So when you're using this kind of core, companies can charge customers to run an open source model on their cloud servers.
It can be charged for your fee, and then they can also provide technical support
Fine-tune the model for a particular company's needs or offer platforms that allow businesses to build their own AI agents and applications, like the example that you just listed, Fei-Fei.
This is probably a city or local government platform.
They don't have a technician to develop their own chatbot.
And then they're using...
The core of DeepSeek and they may hire somebody to maintain this platform.
So this will all be charged by these companies.
They call it, they don't download it.
And there are also subscriptions and paid tools because for Moonshot, Kimi K3, I am a paid user of that model actually.
And they also launched like different layers, different hierarchies of payment that you can pay.
So that you can, I think for me, I pay the bare minimum of paid subscription on Kimi 3.
So when they are asking that model to, for example, write a report based on a certain theme, they can start to at least give me a better report compared with the free version of it.
Yeah, I guess that's a good way to put it.
It seems like you get a smarter version of itself.
Yeah, yeah.
And also, when it comes to if there are a lot of users, too many users are using the same model at the same time.
I think we talked about that during that when we were explaining what token is.
You can also sort of have privilege to use the token that you need compared with people, for example, don't pay or subscribe to this model.
The model itself is like a free front door of the restaurant that we just listed.
Once people walking,
Yeah.
Yeah.
So it's maybe if you buy a piece of gym equipment that only works at one gym.
I don't think that's a great analogy, but that just popped into my mind because it's not forcing customers to come back to the creator of the model, but it does make a lot of sense a lot of the time.
Mm-hmm.
All right.
Now let's talk about that.
That's from a business point of view.
What about people like you and I?
What difference does it make if we have more companies that are open source now, if we have a lot of the open source downloads, you know, coming from wonderful Chinese companies?
What difference does this make for us?
I think from an ordinary person, a citizen's point of view, what I can see from this year is that more AI chatbots or AI platforms are just surrounding us in our daily lives.
Basically, it feels like every...
No matter what kind of feature they have or what kind of area industry they are in, they're offering you an AI solution.
Somehow, someway.
Right?
Which means that, you know,
The most important change is that open source AI is increasingly being connected to the real economy.
And according to China's Ministry of Industry and Information Technology,
The country's AI plus manufacturing initiative aims to promote 500 typical application scenarios by 2027.
And that means AI isn't only being imagined as something that answers questions and maybe writes emails, it is increasingly being connected to applications.
A lot of areas, manufacturing, energy, transportation, healthcare, and other industries.
Yeah, healthcare.
There's a hospital in Jiangsu that now uses a free adapted AI to help doctors check for symptoms and spot drug risks and remind patients about medications and things like that.
And we talked about the use of AI in hospitals, right?
Well, Jiangsu, I'm not sure if they used the open sourced model from DeepSeek or if it was another open source model, but...
And this is where it makes a difference for hospital patients and for doctors too, is that that hospital,
If they didn't have access, you should correct me anywhere if I'm wrong here, but this is how I understand it.
If that hospital didn't have access to an open source AI model,
Then they would have to completely build their own AI model from scratch.
Now, those are doctors and nurses, and I don't know if they know how to build an AI model from scratch, but probably not.
And also when it comes to the payment, if you want to build an AI model from scratch for a small hospital in a county in a Chinese province, I don't think they have the resources to start.
So they take that AI model and they have their own team?
Or I'm guessing they can hire people from the original creator of the open, the AI open source, like you just talked about before, Yuxin, from the business model.
To help create a personalized platform for that specific hospital in Jiangsu.
And because that hospital in Jiangsu has that model, that AI model in their hospital to help doctors and help patients, that is one of many examples how it affects us in our everyday lives.
It's simply just probably before they may not even have the service or AI service.
And now they're just got a, I think, assistant that connects more.
Between the patients and the doctors because these kind of adapted AI, they can help doctors to check symptoms, to spot drug risks, to remind patients about medications.
Probably compared to last year when a small town doctor had to rely only on memory and also paper files.
Now the doctor has a smart assistant that never gets tired.
What an upgrade.
Yeah.
And another part I think is also a way about digital inclusion.
Because when we are, for example, when we're talking about extreme weather events, when we're talking about natural disasters,
To put it on a global context, for a lot of countries around the world, they don't have the resources to pinpoint every forecast, every natural disasters and extreme weather events because they don't have the resources, don't have satellites.
Of their own.
They don't know they have the platform to make the calculation and forecasts.
And I think in this year in China, they also launched a sort of open source meteorological early warning platform called MAZU.
Which is helping a lot of countries, meaning they put a similar structure to make forecasts, collect data, make early warnings.
And then countries, for example, Pakistan or Djibouti or Mongolia, they can build and tweak data.
That Mazu platform and make their own weather forecasts.
And Mazu, I know it's the local ocean goddess to protect people from the rain.
Yeah, it's a traditional Chinese sea goddess.
Yeah.
And now we are seeing a lot of less developed, least developed countries around the world, like in Africa, Djibouti.
They're also developing their own version of MAZOO that can help improve their forecasting weathers.
And also in Jordan, they are developing their own Arab language model for floods, droughts, and heat waves.
So I think in this way, you're sort of closing that gap for a lot of countries and groups in the society when it comes to this development of AI.
And it sounds like a very advanced and also very expensive technology.
Yeah, it creates a more level playing field.
We hear the leaders of the tech giants talk about open source all the time and how they feel about it.
And the CEO of NVIDIA, Jensen Huang, he said, I think he talked about the fact that there should be a balance between closed models and open models.
The Microsoft CEO said that open source is a key component to having a
Healthy AI ecosystem.
The Google CEO said that I think Google specifically has benefited from open source.
Why?
Why do they say that?
Either having a balanced system between closed and open or open benefits and Google also benefited.
Why are these biggest tech companies in the world benefiting?
Because I would assume Google, you can't make your own stuff.
They're using open source stuff too.
So why is that?
Yeah, I'm really glad that we're talking about this.
This is kind of a deeper layer of why this world is kind of embracing the open source kind of thing.
And of course, open source is widely used in a lot of these technology companies, and there is a practical business reason behind it.
This support and these companies are not just watching the open source system.
They're actually using it, also contributing it.
And NVIDIA releases open models for developers to build AI applications and robotics and other systems on NVIDIA hardware.
The gym equipment that only works at the gym.
Exactly.
At one gym.
So you can understand if the world is widely using open source, of course, the world is using more chips.
And then, of course, NVIDIA is...
Benefiting from this.
And Microsoft lets customers bring different open models into their Azure AI and co-pilot ecosystem.
And Google has released open source tools such as Gemini CL1.
And of course, Android is a very well-known open sourced tool for
The worldwide developers.
So they can build its AI products around a wider ecosystem of open source frameworks.
So open source can expand the entire AI ecosystem and these companies make money from different parts of that ecosystem, no matter it's the software, it's the hardware,
Cloud computing developer tools or AI services.
But not everybody's on board, though, because I already said that Anthropic and OpenAI, there was an article in the New York Times that was talking about this.
Those two companies, OpenAI and Anthropic, they're expressing real concern about this.
They're actually lobbying regulators in Washington to restrict open source models.
The head of public policy, I think her name is Sarah Heck from Anthropic, said that these companies...
What they say is that open source models could be misappropriated.
They could be used for adversarial purposes, intelligence purposes, making them too dangerous for unrestricted use.
And unrestricted, you said the term maybe 15 minutes ago, Yushan, the weights.
The weights are removed.
And correct me again if I'm wrong, when the weights are removed from the AI source, it means it's completely unrestricted.
So everybody can sort of tap in and use it.
And I think that is one of the biggest arguments.
That's why Anthropic and OpenAI say that this could potentially be a little bit dangerous if put in the wrong hands.
I mean, yeah, from their perspective, it is understandable, right?
It's like you are using another company's recipe or maybe the skeleton of that whole robot.
And then you will be concerned about the safety and public security, that kind of thing.
Yeah, I think the robots are...
Pretty easy to understand example, because if you have a robot that's basically like a blank canvas to your engineers, they could do some bad stuff with that.
Then another of their concern could be about intellectual property.
Like they spend a lot of resources in building the skeleton of this robot.
And now somebody else is building more new fleshier things based on their skeleton.
It just feels like that...
Their fruits are being stolen by somebody else, which I think on some level for some of the biggest for OpenAI and now ThrillPrick, they're probably one of the biggest AI companies in the world.
Yeah.
And the AI, it's getting more powerful.
It's getting smarter.
It's getting faster.
And I think what they're saying is what it's capable of in the future.
You know, some of these technologies are too powerful to release without some safeguards on what they are capable of doing.
All right.
Well, let's leave it there.
I know that we talked about this for quite a while today, but on Roundtable, what we like to do is take a topic that's maybe not so easy to understand if you go to Layer 1, Layer 2, Layer 3 and beyond and make it accessible for everyone.
Including myself.
And I feel, how do you feel, Fei-Fei?
I feel like I know more today.
Yeah, I learned a lot.
Keywords, parameter, weights, so on and so forth.
Open source, closed source.
I now have a clearer picture.
And Yuxin put this together, this explanation, lengthy explanation together for us.
So thanks for that, Yuxin.
Very good.
