The Future of Qwen: A Generalist Agent Model — Junyang Lin, Alibaba Qwen

Hey everyone, I'm Jun-young from Quentin in Alibaba Group. I'm very happy to be here in the AI Engineer World Fair 2025, and I'm very excited to share some progress about Quen the auto view. I think you guys know about Quen, and maybe you are developers.

I'm very happy to share some more things about Quen to all of you. Quen is a series of large language models and large multimodal models, and we have a dream of building a generalist model and generalist agent. And before I start, I would like to share some important links to all of you.

Maybe you know our product, our chat interface, Quen Chat, which is chat.quen.ai. It is very easy to use. You can interact with our latest models, you can use our multimodal models by uploading images and videos, and you can also interact with our omni models using voice chat and video chat.

And there are some important features like web dev and deep research. Welcome to enjoy it. And if you'd like to know more about technical details, you can check our blog, which is quenlm.github.io. In our blog, once we release something new, we often release a blog, so you can know more about technical details through our blogs.

As we keep open sourcing, so we have our codes in GitHub, and we have our checkpoints in Hugging Face, so you can download our checkpoints and play with our models. Yeah, so feel free to check our websites and enjoy them.

Yep. In this year, just before the Spring Festival, we have released a very good instruction tuned models. I think it's a very good basis for large language models, which is Quen 2.5 Max. It is a very large MOE models, and we find that in multiple benchmarks, it achieves very, very competitive performance to the state-of-the-art models at that time, including Cola 3.5, GPT-4, and Deep CPT-3.

But

we believe that there are more potential for a large language model, not just becoming an instruction tuned model, but it can be smarter and smarter with reinforcement learning. So we have dived into doing the research in reinforcement learning, and we finally found that, well, it is really amazing to see that using RL can increase its performance, especially in reasoning tasks like math and coding, and increase the performance quite consistently, just like for AIME 2024.

For a 32 billion parameter model, you can find that its performance starting from like 65 and increasing until it is 80. So it is really amazing to build a reasoning model like QWQ. And also in Chapel Arena, you can find that its performance is also very competitive with even larger models, and it is in top 15 for a long time.

So we would like to combine all our efforts in our research and development to build a stronger next generation models. So very, very recently, we have released Quen3. Quen3 is our latest large language models, and we have released multiple sizes of dense and MOE models, a large number of models.

First of all, I would like to share the flagship model to all of you, which is a

235 billion parameter model, with a total number, but only activates 22 billion parameter models. It is an MOE models, so it is actually efficient, but it's very effective in comparison with very top-tier models like O3 mini, and it's just lagging a little bit behind Gemini 2.5 Pro.

And also for our largest dense model, it is also very competitive as well. And we have a very, very fast MOE models. It is relatively small, although it has 30 billion total parameter models, only activate 3 billion parameter models.

But if you compare the performance with QWQ, 32 billion, the MOE models, which only activate 3 billion parameter models, even can outcompete QWQ, 32 billion in some tasks. And for an even smaller models, very small, it is only 4 billion parameter models.

But this time we have a lot of techniques in distillation, distilling the knowledge from large models to small models. We finally built a very good small models, and with thinking capabilities, it even shows competitiveness to the flagship model in our last iteration, like Quen 2.5 72B.

Yeah, so it is a really good model, and it is really worth a try for you to play with our 4 billion parameter model. And the 4 billion parameter model, and you can even deploy it in mobile devices.

Yeah. So for Quen3, we have some important features, and the most important features this time, it is hybrid thinking mode. So what is hybrid thinking mode? Hybrid thinking mode means that you can use thinking and non-thinking in a single model.

We combine the two behaviors into a model. What is thinking mode? Thinking is like, before you answer the question with a detailed answer, you just start thinking. You just reflect on yourself, explore possibilities. Finally, you find that, well, you are ready to answer the question, so you provide the answer.

So like O1 and DeepSeek R1, they have the thinking behavior. Yeah. And for non-thinking mode, it is just the traditional instruction tuned model. It's just like a chatbot without thinking, without lagging. It is just provide the answer. It is near instant.

But this might be the first time in the open source community that we combine the two modes into a single model. So you can use prompts or hyperparameters to control its behavior as you like. And once we dive into the hybrid thinking modes, we find that we can create a feature of dynamic thinking budget.

So what is dynamic thinking budgets? For thinking budgets, which means the maximum thinking tokens. For example, if you have 32,000 tokens for your thinking budgets, so for a task which requires thinking, for example, if you finish the thinking with like 8,000 tokens, it is below 32,000.

Okay, fine. You finish thinking, provide the answer. It is good. But if you only have a thinking budget of 4,000 tokens, so if your thinking requires more than 4,000 tokens, like you require 8,000 tokens to finish your thinking, but you only have a budget of 4,000 tokens, so you will stop at 4,000 tokens, which means that your thinking process is truncated.

So we can check the performance with larger and larger thinking budgets, and you'll find that, well, the performance increase quite well with the increase of thinking budgets. And with a very small thinking budget, for example, in AIME 24, it only achieves just over 40.

But if you have a large thinking budget, like 32,000 tokens, you can even achieve more than 80. It is really, really amazing with the thinking capabilities. So I hope you enjoy the hybrid thinking mode. You only single model to achieve thinking and non-thinking, and you'll find something good for you.

For example, in your task, you require only like 95% accuracy, and you will find that, well, for example, you use only 8,000 tokens for your thinking budget, you can achieve over 95%. So that is quite well. You don't need to waste more tokens on your thinking.

So you can keep your thinking budgets with 8,000 tokens. This is only an example. We would like to explore the usages. Yeah. The next important features is that Quen3 supports over 119 languages and dialects. In Quen 2.5, it only supports 29 languages, but this time we support over 119 languages and dialects.

We have detailed names of the languages and dialects that we support. You can check it. I think it would be really good for global applications, and there are a lot of people previously, especially if they are using open weight models.

The open weight models don't support many languages quite well, so there will be more people that are capable of using large language models in their domains and in their languages. Yeah. And we have specifically increased the capabilities in agents and codings, and especially we enhanced the support for MCP, which is really popular very recently.

And these are two examples to show that our models how to use tools during thinkings. So you'll find that it can think while it is using a function calls to use the tools, and it gets the feedbacks from the environment, and it keeps thinking.

So that would be a feature that we really prefer, which is the model is capable of thinking, but it is also capable of interacting with the environment, and it keeps thinking. It is really good for like inference time scaling.

And this is another example for the model to organize the desktop. So if it has the access to the file system, it can do things like that. It thinks which tools should it use, and then it uses the tools and gets the feedback and continues thinking while it finishes using the tools, and it finishes the task and tells you that, well, we have organized the desktop quite well.

And these are two very simple examples to show that we have provided better and better support for agent capabilities. And we not only would like our models to be a simple chatbot, we would like it to be really productive in our working life, to become a really, really productive agent.

Yeah. And these are three features of

Quen3. Yeah. We have open weighted a lot of sizes, including two MOE models. A small MOE model has a total number of 30 billion and only activate 3 billion, but another one is 235 billion, activates

22 billion parameter models. Yeah. We also have six dense models, and for smaller models, you can use it for testing, and you can support draft models, but for 4 billion parameter models, you can deploy it on mobile devices.

And for 32 billion parameter models, these are models that you really prefer, and especially 32 billion parameter model, it is strong. It shows competitiveness, and you can use it for doing RL. You can deploy it in your local environment as well.

So we also open weighted dense models as well. But we believe that in this year, maybe in the next years, the future trend is belonging to the MOE models. So later we will release more MOE models for you to use, and there will be better support in the open source community, like third-party frameworks for the MOE models as well.

Yeah. Besides building large language models, we are also building multimodal models. We have focused quite a lot on vision language models, and I think many of you maybe are using Quen2.5-VL and now using Quen2.5-VL, which was released this year in January.

It achieves very competitive performance in vision language benchmarks, like understanding benchmarks like MMU, and also benchmarks like MathVista, and also a lot of general VQA benchmarks. So it achieved very good performance in the benchmarks for vision language understanding.

We also explore the capabilities of thinking for vision language models. So we have built QWQ as well, and we find inference time scaling with larger maximum thinking length, which is equivalent to the thinking budget that I talked before.

So if you have a larger thinking budget, it will achieve better and better performance in reasoning tasks, especially like mathematics. Even for vision language models, it shows similar features as well. But for multimodal models, what we really would like to do is to build an omni model, which accepts multimodalities for the input, and also it is capable of generating multiple modalities, like text, vision, and audio.

But this time, this is not the perfect state, but I think it's good. It's relatively a small model, but we are really proud of this attempt. It's a 7 billion large language model is based on that, but it is capable of accepting three modalities, including text, vision, vision includes images and videos, and also it can accept audio, and this time it can generate text and audio.

Maybe in the future, our models might be capable of generating images, high-quality images and videos. That will be a truly omni models. Now for this omni model, it can be used in voice chat and video chat and text chat as well.

Yeah. It achieves similar performance in audio tasks for the same size models, like 7 billion parameter models. But what surprised us a little bit is that it can even achieve better performance in vision language understanding tasks in comparison with Quen2.5-VL's 7 billion, which means that we can achieve very good performance for an omni model in vision language tasks.

But a little bit that we have not done well, but we believe we can do well is that we should recover the performance drop in language tasks, especially for its intelligence, especially for its agent tasks. I think we can recover it by improving our data quality, improving our training methods.

But now there is still some room for improving the model capabilities in different domains and different tasks. And this is the omni models. And no matter what models, we keep open sourcing. We love open sourcing because open sourcing really helps us quite a lot

by the developers can give us some feedbacks to help us improve our models. The interaction with the open source community makes us happy and makes us we have more we are more encouraged to build more good models for all of you.

Yeah. We have a lot of titled models in the open source communities, including LLMs and also coders. And Quen2.5 coders is something that

many people are using for local development. And something that I can tell you that we are now building Quen3 coders. I think you guys know about it. Yeah. We have many model sizes because we just believe that for each size, there might be a lot of users.

And actually, there are. There are a lot of users, no matter very, very small models, like 0.6 billion parameter models. And very large models, previously 72 billion dense models, and now 235 billion MOE models. There are a lot of users using it.

And they need quantized models. So we just provide quantized models in different formats, including GGUF, GPQ, AWQ, and MLX for Apple as well. We try to use Apache 2.0 for most models. So you can just use it freely.

You can change the models freely in your business. You don't need to worry about too many things. You don't need to access for the permission. You just directly use it. We hope that large language models and large multimodal models or foundation models can help you create good applications.

And this is what we like to do. And as we are becoming popular and popular during these two years, so for Quen models, it should be supported by a lot of maybe most relative third-party frameworks and API platforms as well.

Yeah. And we are also building products. We are also building products for you to interact with our models. We are building agents as well. And in our Quen chat, just as I mentioned before, we have some very good features.

This is something that I really like, which is called Web Dev. By entering Web Dev, you just need to insert or input a very simple prompt, things like create a Twitter website. And you'll find that it just generates a code and then some effects like artifacts.

So you have a website to see how it is shown. And you can also deploy it as well. You can deploy it and get the URL, share it to your friends to show that how creative you are. You can also create a website for your product.

For example, this is a very simple prompt. Create a sunscreen product introduction website. And you have a very good website. You can even click on the buttons as well. And for making a card, not just making a website, just by making a card, well, it generates a card.

I really like the card. We often use the card in our Twitter. So our prompt is also simple. We give it a link. So based on the link, you just create a nice looking card, and we provide more information to you.

You just based on the information and build a card for us. Yeah. This is something I really, really like. Web Dev makes me more creative and helps me quite a lot to show our things to people all over the world.

Yeah. We also have things like deep research. By doing deep research, you just need to access something to write a report in what you are interested in. Yeah. Maybe like healthcare industry, maybe like artificial intelligence. Just access it.

Give it a prompt. And it will ask you what you are going to focus on. You can tell it. Or you can just say as you like. And it will start doing research by making a plan first and then doing the search step by step, writing parts step by step.

Yeah. And keep searching and finally gives a comprehensive report to all of you. You can download our report by a PDF. We are still improving its quality by doing reinforcement learning to build a fine-tuned model specifically for deep research.

And we believe that there's still much room in this field. It is really hard to do reinforcement learning for this stuff, but once you have built a good model for this product, it will be really productive for people in their working life.

Yeah. So in the future, we will do many things. There are still a lot of things for us to do to achieve AGI, to build a really good foundation model and foundation agent for all of you. And the first thing is that something really different from what other people think.

We still believe that there is still much room in training. I'm happy that you have shown your preference in our pre-training models, but we still find that there are still a lot of good data we did not put into it.

There are still a lot of data we did not clean it quite well. And we find that we can use multimodal data and to make the models more capable in doing different tasks in different domains. And we also have synthetic data.

And maybe we will finally do something really, really different training methods for pre-training, not just like next token prediction. Maybe later we use reinforcement learning in pre-training as well. And so there is still much room in pre-training to build a very good basis for the chatbot or agents.

And this is the first thing. And the second thing is that the scaling laws, there are some changes in scaling law. Previously, we were scaling in the model sizes in the pre-training data, but now we need to scale the compute in reinforcement learning.

And we have a focus on long horizon reasoning with the environment feedback. So if you train the model, which is capable of interacting with the environment, keep thinking, this will be something really, really competitive. So it will get the feedback from the environment, keep thinking.

It will become smarter and smarter with inference time scaling. Yeah. So you'll find that it will generate very, very long context. And you have very, very long context for your input, especially when you have memory. So you need to scale your context.

Maybe finally we are moving towards infinite context. But now we need to fix the problems of 1 million tokens quite well. And then we are marching to 10 million tokens and then infinite context. So we are scaling context.

We are going to scale the context at least 1 million tokens this year for most of our models. Yeah. We're also going to scale modalities. Maybe scaling your modalities doesn't increase your intelligence, but if you scale your modalities, you can make your models more capable and more productive, especially with the vision language understanding.

If you have vision language understanding capability, you can make a GUI agent. But before that, if you have no vision capability, you can. It is almost impossible for you to make a GUI agent and do things like computer use.

And maybe there is still much room in scaling the modalities in either inputs and outputs. We are going to unifying understanding and generation. For example, for the image understanding and image generation at the same time, just like GPT-4.0, they generate very interesting and high-quality images.

That is something what we are going to do as well. Yeah. So based on the four things that I mentioned, so if you would like me to summarize what we are going to do in this year and next year, I think we are moving from the era of training models to training agents.

We are actually training the agents, not only with scaling with pre-training, but also scaling with RL, especially with the environment. We are actually training the agents. So I think we can say that we are now staying in the era of agents.

And that's all. Thank you very much for listening to my talk. And if you are interested in Quen, shoot me an email and talk to me in X. Yeah. Thanks a lot. Yeah.