AI Engineer World's Fair 2025
The Future of Qwen: A Generalist Agent Model
Read the talk
Qwen’s path from language models to generalist agents
Hybrid thinking, tool feedback and multimodal inputs turn model capability into agent behavior, while Qwen’s roadmap expands the compute, context and training needed to sustain it.
From a talk by Junyang Lin
Before you start: Basic familiarity with language models, tokens and tool calling is helpful; the article explains mixture-of-experts models and thinking budgets as they arise.
What does a generalist model need to do?
How can one model family move beyond answering text questions to working with images, video, speech and tools? Qwen starts with both language and multimodal models, with the longer-term goal of building a generalist agent. In Junyang Lin’s 2025 presentation, Qwen Chat is the entry point: users can upload images and videos, interact through voice or video chat, and try applications such as WebDev and deep research.
For developers, the Qwen release blog provides technical explanations, while GitHub code and Hugging Face checkpoints provide a route from trying a hosted model to running and adapting one. Those access points matter because the family spans different tasks and deployment sizes; there is no single checkpoint that represents every Qwen capability.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From instruction following to reasoning
Qwen2.5-Max, released before Spring Festival 2025, establishes the instruction-following baseline. It is a large mixture-of-experts, or MoE, model. Lin describes its benchmark performance as competitive with Claude 3.5, GPT-4o and DeepSeek-V3. The accompanying chart compares models across Arena-Hard, MMLU-Pro, GPQA-Diamond, LiveCodeBench and LiveBench, separating several kinds of capability rather than reducing them to one score.
The next question is whether reinforcement learning can make a model better at solving problems, rather than merely following instructions. Lin reports consistent gains in mathematics and coding while developing QwQ-32B. Lin reports that a 32-billion-parameter model’s AIME 2024 score rose from roughly 65 to 80 during reinforcement learning. The presentation does not specify the checkpoint sequence or scoring protocol behind that trajectory. He also reports sustained top-15 placement in an arena evaluation and competitiveness with larger models, without establishing a usable leaderboard identity for that comparison.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Capability at different compute scales
Qwen3 combines the instruction-following and reasoning work in a family of dense and MoE models. The flagship, Qwen3-235B-A22B, has 235 billion total parameters but activates 22 billion per token. This distinction is central to MoE efficiency: total capacity and the parameters used for a particular token are different quantities. Lin places the flagship near o3-mini and slightly behind Gemini 2.5 Pro in the comparisons he presents. He also describes the largest dense model as competitive.
The smaller models make two different efficiency arguments:
- Sparse activation: Qwen3-30B-A3B has 30 billion total parameters and activates 3 billion. Lin says it can outperform QwQ-32B on some tasks despite activating far fewer parameters.
- Distillation: Qwen3-4B transfers knowledge from larger models into a smaller dense model. With thinking enabled, Lin describes it as competitive with the previous Qwen2.5-72B flagship and suitable for mobile deployment.
These are different routes to reducing inference requirements: MoE selects a subset of a larger model, while distillation improves a smaller model’s capabilities. The mobile claim comes without a device, quantization setting or latency measurement.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Choosing when to think
Hybrid thinking puts two response behaviors in one model. Thinking mode explores possibilities and reflects before producing the answer, as in o1 and DeepSeek-R1. Non-thinking mode behaves like a conventional instruction-tuned chatbot, moving directly to the response with less delay. The useful control is whether a request needs deliberation, rather than whether the application should switch to an entirely different model.
Prompts or configuration parameters select the behavior. Lin tentatively presents combining both modes as a possible first for the open-source community; the practical feature is the shared model, regardless of priority. The original Qwen3 release exposes enable_thinking through its Python chat template. For example, an application can prepare the same request in either mode:
python
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
messages = [
{"role": "user", "content": "What is 17 × 24?"}
]
prompts = {
mode: tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=enabled,
)
for mode, enabled in {
"thinking": True,
"non_thinking": False,
}.items()
}
print(prompts["non_thinking"])
This selects the prompt format; it does not itself run generation or set a thinking-token budget.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
How much thinking is enough?
Once a model can think before answering, the application needs a way to limit that work. A thinking budget is the maximum number of tokens available for reasoning. Consider Lin’s example of a task whose reasoning would finish after 8,000 tokens:
| Thinking budget | What happens |
|---|---|
| 32,000 tokens | Reasoning finishes after 8,000 tokens; the model can answer. |
| 4,000 tokens | Reasoning reaches the cap before completion and is truncated. |
A budget is a ceiling, not a requirement to consume every token. It also differs from a total output limit, which can constrain the answer as well as reasoning. Current implementation guidance separates capped reasoning from final-answer generation; the presentation does not specify its exact implementation.
Lin reports AIME 2024 performance just above 40 with a very small thinking budget and above 80 with a 32,000-token budget. The smallest budget, model variant and sampling protocol are not specified in the passage. The slide places AIME24 alongside AIME25, LiveCodeBench and GPQA Diamond, showing thinking-mode curves against non-thinking baselines. The engineering question is how much additional reasoning improves the task enough to justify its token cost.
Lin illustrates that decision with a hypothetical application requiring 95% accuracy: if an 8,000-token budget already exceeds the requirement, there is no need to allocate more thinking tokens merely because the model permits it. The example is a budget-selection rule, not a measured 95% result. He then turns from reasoning controls to another part of the release: broader language coverage.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From multilingual answers to tool interaction
Qwen3’s release specifies support for 119 languages and dialects, compared with 29 languages for Qwen2.5. The release provides the detailed language list. The motivation is practical: developers working in languages underserved by earlier open-weight models need models they can use in their own domains, not just models that perform well in a few widely represented languages.
The next capability changes the structure of a response. Qwen3 strengthens agent and coding support, including Model Context Protocol (MCP) integration. A function call can occur during thinking; the tool’s result then becomes new evidence for continued reasoning. Instead of generating an entire plan from a fixed prompt, the model can revise its next step after observing what happened.
Desktop organization makes that loop concrete. Given filesystem access, the model must select tools, use them, inspect their feedback and continue until the work is complete:
- Determine which filesystem tool is needed for the next step.
- Invoke the tool and receive its result.
- Continue reasoning with that environmental feedback.
- Repeat as needed, then report the completed organization to the user.
The desktop example is one of two tool-use examples Lin introduces. Its distinctive feature is thinking that continues through interaction: tool use is part of the reasoning process, and the environment helps determine what happens next. That is the bridge from a chatbot to an agent doing useful work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A lineup for local experimentation
The initial open-weight Qwen3 lineup contains two MoE models—the 30B/3B-active and 235B/22B-active variants—and six dense models. The sizes serve different deployment and experimentation roles:
| Model group | Uses highlighted by Lin |
|---|---|
| Small dense models | Testing and draft-model roles |
| 4B dense | Mobile deployment |
| 32B dense | Local deployment and reinforcement-learning experiments |
| MoE models | Larger total capacity with sparse activation |
Lin highlights Qwen3-32B as a strong model developers can run locally and use for RL. Dense models remain part of the release, but he expects future development to favor MoE, with more releases and better support from third-party frameworks.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Reasoning over visual inputs
The same progression extends to vision-language models. Qwen2-VL was followed by Qwen2.5-VL in January 2025. Lin cites competitive performance on MMMU, MathVista and general visual question-answering benchmarks. These evaluations cover understanding and reasoning about visual inputs, rather than only generating text from text.
QVQ explores thinking in this setting. Increasing the maximum thinking length gives the model more room to reason about visual problems, especially mathematics. The displayed MathVision chart reports accuracy rising from 43.5% at 4,000 thinking tokens to 48.1% at 24,000, with intermediate points of 45.6% at 8,000 and 46.7% at 16,000. The exact QVQ variant and evaluation protocol are not identified here, but the slide makes the proposed scaling axis explicit: more inference-time reasoning for a visual task.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Combining modalities without losing capability
An omni model broadens both what the model can perceive and what it can produce. Qwen2.5-Omni is built on a 7B language model and accepts text, vision—including images and video—and audio. Its outputs in this release are text and audio, enabling text, voice and video chat. High-quality image and video generation remain future ambitions, not capabilities of this demonstrated release.
Lin reports leading audio-task performance among similarly sized 7B models and is encouraged by the visual results. The comparison with Qwen2.5-VL-7B is task-dependent: the original technical report shows Omni ahead on MMMU validation but slightly behind on MathVista testmini. Combining modalities can therefore preserve strong visual understanding without uniformly outperforming the specialist vision model.
There is also an acknowledged regression. Language intelligence and agent-task performance have dropped, and Lin proposes better data quality and training methods as ways to recover them. Broader modality coverage does not automatically preserve every existing capability. The goal is a model that can work across domains without sacrificing the reasoning and agent behavior that make those inputs useful.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Making the models usable
Open releases create a feedback loop with developers: people use the models in different settings, report what fails and help guide improvements. Qwen2.5-Coder is already used for local development in Lin’s account, while Qwen3-Coder is still under development at the time of the talk. The breadth of the lineup reflects demand at every scale, from 0.6B models to the earlier 72B dense models and the 235B MoE.
Distribution also means supplying formats people can deploy. Lin lists GGUF, GPTQ, AWQ and MLX for Apple hardware. He says most Qwen models use Apache 2.0, allowing business use and modification without separately requesting permission; that family-level statement should not be read as a license declaration for every checkpoint. The initial eight Qwen3 models are explicitly Apache 2.0. Broad third-party framework and API-platform support is another part of making the family accessible beyond its own chat interface.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
WebDev: a prompt becomes a shareable artifact
Qwen Chat packages model capabilities into workflows. In WebDev, a simple request to create a Twitter website produces code and an artifact-like preview. The next step is deployment: the user can obtain a URL and share the resulting site. This connects generation to an inspectable output rather than leaving the user with code alone.
A sunscreen product introduction site provides another example, including clickable buttons. The same workflow can produce something smaller than a website: a card based on a supplied link and additional information. Lin says the team uses these cards on Twitter. Across the examples, the input stays simple while the output becomes a visual object the user can inspect, interact with or share.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Deep research: planning, searching and writing
Deep research turns a broad topic into a longer sequence of work. A user might request a report about healthcare or artificial intelligence. The system first asks what to focus on; the user can narrow the scope or leave that choice to the system. The workflow then proceeds through several stages:
- Make a research plan.
- Search step by step.
- Write parts of the report incrementally.
- Continue searching as the report develops.
- Assemble a comprehensive report that can be downloaded as a PDF.
Search and writing are interleaved rather than treated as one retrieval followed by one final response.
Lin describes reinforcement-learning fine-tuning for a specialized deep-research model as ongoing and difficult. A usable research product requires sustained progress across a sequence of decisions, not just a good answer to one prompt. There remains substantial room to improve the quality of that process, but its attraction is concrete: reducing the work needed to produce a useful report.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Scaling the training and environment of an agent
The future agenda begins with pretraining, where Lin still sees considerable headroom. Some useful data has not been included; other data has not been cleaned well enough. Multimodal and synthetic data could expand the tasks and domains a model handles. He also raises the possibility of changing the training objective beyond next-token prediction, including reinforcement learning during pretraining. That last proposal is an avenue to explore, not a completed training recipe.
The second direction is scaling reinforcement-learning compute. Earlier scaling emphasized model size and pretraining data. Agent training adds long-horizon reasoning with environmental feedback: act, observe the result, continue thinking and decide what to do next. This returns to the desktop-organizing mechanism, but now as a training objective. Inference-time scaling can include repeated interaction with the world, not just a longer uninterrupted stream of reasoning.
That interaction creates the third pressure: context. Long reasoning trajectories generate more history, while memory increases the input an agent needs to use. Lin’s proposed sequence is to handle one million tokens well, then pursue ten million and eventually an aspirational infinite context. Lin sets a 2025 target of at least one million context tokens for most Qwen models. This is a roadmap commitment; the original Qwen3 release lists 32K or 128K context depending on the model.
The fourth direction is input and output modalities. Lin distinguishes increased productivity from increased intelligence: another modality can enable a new kind of work even if it does not make the model a better reasoner. Vision-language understanding, for example, enables a GUI agent to inspect an interface and perform computer-use tasks; he describes such work as almost impossible without vision. The output side matters too. Unifying image understanding with image generation would let one system both interpret and create visual material, with GPT-4o’s image generation serving as the example.
The development target is shifting from training models to training agents. Pretraining supplies the foundation, while reinforcement learning—especially interaction with environments—trains behavior across extended tasks. Larger contexts and more modalities give those behaviors room to operate. For Lin, the agenda for 2025 and the following year is to bring these directions together so that the model can keep working, observing and reasoning until a task is done.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The public Qwen chat interface, now presented as Qwen Studio.
Official weights and usage guidance for the 32-billion-parameter reasoning model.
Further reading
The original Qwen3 release overview, with model sizes, language coverage, hybrid thinking controls and deployment examples.
The Qwen team's account of outcome-based reinforcement learning for mathematical reasoning, coding and general capabilities.
- Qwen2.5-Max release overviewArticle
Historical benchmark comparisons and training background for the large MoE model.
- Qwen2.5-VL release overviewArticle
Vision-language capabilities spanning documents, video, localization and computer use.
The December 2024 introduction to QVQ's experimental visual reasoning.
The original technical report, including task-specific comparisons with specialist vision and audio models.
Qwen's agent framework with tool calling, MCP integration and code-interpreter support.
Updates since the talk
Current example separating budget-limited reasoning from final-answer generation.
Read the complete timestamped transcript
- 0:03
Hey, everyone. I'm Junyang from Qwen team, Alibaba Group. Very happy to be here in AI Engineer World's Fair 2025, and I'm very excited to share some progress about Qwen to all of you.
- 0:14
I think you guys know about Qwen and, uh, maybe you are developers. I'm very happy to share some more things about Qwen to all of you. Qwen is a lar...
- 0:23
a series of large language models and large multi-model models and, uh, 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.
- 0:35
Uh, maybe you know our product, our chat interface, Qwen Chat, uh, which is, uh, chat.qwen.ai. Uh, it is very easy to use. You can interact with our latest models.
- 0:45
You can use our multi-model models, uh, 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, um, deep research.
- 0:59
Welcome to enjoy it. And, uh, if you'd like to know more about technical details, you can check our blog, which is qwenlm.github.io. In our blog, uh, once we release something new, we often, uh, release a blog, so you would...
- 1:13
can know more about technical details, uh, through our blogs. Uh, as we keep open sourcing, so, uh, we have our codes in GitHub, and, uh, we have our checkpoints in, in Hugging Face, so you can download our checkpoints and play with our models.
- 1:28
Yeah. Um, so feel free to, uh, check our websites and enjoy them. Yep.
- 1:35
Um, in this year, just before the Spring Festival, we have released a very good instruction tune models. I think, uh, it's a very good basis, uh, for large language models, which is Qwen two point five Max.
- 1:49
It is a very large MoE models, and we find that, um, in multiple benchmarks, uh, it achieves very, very competitive performance to the state-of-the-art models at that time, including Claude 3.5, GPT-4o, and DeepSeek-V3.
- 2:04
Uh, but we, uh, we, uh, we believe that there are more potential for a large language model, not just becoming an instruction tune model, but it can, uh, be smarter and smarter with reinforcement learning.
- 2:19
So we have dived into, uh, doing the research in reinforcement learning, and we finally found that, well, it is really amazing to see that u-using RL can increase its performance, especially in reasoning tasks like math and coding, uh, in-increase the performance quite consistently, uh, just like for, for AME twenty, uh, twenty-four.
- 2:41
For a thirty-two billion parameter model, you, you can find that its performance, uh, starting from like sixty-five and is in-increasing until it is eighty. So it is really amazing to build a reasoning model like Q, uh, QWQ.
- 2:56
And also in Treble Arena, you can find that its performance, uh, it's also very competitive with even larger models, and it is in, uh, top fi- uh, fifteen for, for a long time.
- 3:07
So we would like to combine all our, uh, efforts, uh, in our research, uh, and development to build a stronger next generation models. So very, very recently, we have released Qwen3.
- 3:21
Qwen3 is our latest large language models, and we have released multiple sizes of dense and MoE models and a large number of models. Uh, first of all, I would like to share the...
- 3:33
a flagship model to all of you, which is a, um, twenty, uh, two hundred and thirty-five billion parameter model with a total number, but only activates twenty-two billion parameter models.
- 3:46
Um, uh, it is an MoE models. So it is actually efficient, but it's very effective in comparison with, uh, very top-tier models like o3-mini, and it's just lagging a little bit behind, uh, Gemini two point five Pro.
- 4:03
And also for, for our largest dense model, it is also very competitive as, uh, as well. And we have a very, very fast MoE models. It is relatively small.
- 4:15
Although it has thirty billion total parameter models, on-only activate three billion parameter models. But if you compare the performance with QWQ thirty-two billion, the MoE models, which only activates three billion parameter models, even can outcompete QWQ thirty-two billion in some tasks.
- 4:35
And for a even smaller models, very small, it is only four billion parameter models. But this time we have a lot of techniques, uh, in distillation, distilling large, uh, the knowledge from large models to small models.
- 4:49
We finally built a very good small models, and with thinking capabilities, it even shows competitiveness to the flagship model in our last ver... uh, iteration, like Qwen two point five seventy-two B.
- 5:07
Yeah. So it is a really good model, and it is r-really worth a try for you to play with our four billion parameter model. And the four billion parameter model, and you can even deploy it in mobile devices.
- 5:19
Yeah. So for Qwen3, 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.
- 5:39
We combine the two behaviors into a model. What is thinking mode? Uh, thinking is like, uh, before you answer the question, uh, with an, uh, detailed answer, you just start thinking.
- 5:54
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, uh, like o1 and DeepSeek-R1, they have the thinking behavior.
- 6:08
Yeah. And for non-thinking mode, it is just the traditional, uh, instruction tune model. It's just like a chatbot without thinking, without lagging. It is just provide the answer. It is near instant.
- 6:23
But this might be the first time in the open source community that we combine the two modes, uh, into a single model. So you can use prompts or hyperparameters to control its behavior just as you like.
- 6:36
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.
- 6:51
For example, if you have thirty-two thousand tokens for your thinking budget, so, uh, for a task which requires thinking... For example, if you finish the thinking with, like, eight thousand tokens, it is, um, below thirty-two thousand.
- 7:11
Okay, fine. You finish thinking, provide the answer. It is good. But if you only have a thinking budget of four thousand tokens,
- 7:21
so, um, uh, if, if your thinking requires more than four thousand tokens, like it requires eight thousand tokens to finish your thinking, but you only have a budget of four thousand tokens.
- 7:33
So you will stop at four thousand tokens, which means that your thinking process is truncated. So, uh, we can check the performance with, uh, larger and larger thinking budgets, and you'll find that, well, the performance increased quite well with the increase of thinking budgets.
- 7:50
And, uh, with a very small thinking budget, uh, for example, in AME twenty-four, it only achieved just over forty. But if you have a large thinking budget, like thirty-two thousand tokens, you can even achieve more than eighty.
- 8:04
It is really, really amazing with the thinking capabilities. So, uh, I hope you enjoy the hybrid thinking mode. You, you only single model to achieve thinking and non, non-thinking, and you'll find something, uh, good for you.
- 8:20
For example, uh, in your task, you require only, like, ninety-five, uh, uh, percent accuracy, and you'll find that, well, uh, for example,
- 8:31
y-you use only eight thousand tokens for your thinking budget, you can achieve over ninety-five percent. So that is quite well. You don't need to waste more tokens on your thinking, so you don't...
- 8:41
can keep your thinking budget with eight thousand tokens. This is only an example. We would like to explore the usages. Yeah. The next important features is that Qwen 3 supports over one hundred and n-nineteen languages and dialects.
- 8:58
In Qwen 2.5, it only supports twenty-nine languages, but this time we support over one hundred and nineteen languages and dialects. We have, uh, detailed, uh, names of, uh, the languages and the dialects that we support.
- 9:14
You can check it. I think it will be really good for global applications, and there are a lot of people, uh, previously, especially if they are us-using open-weight model.
- 9:25
The open-weight models don't support many languages quite well, so there will be more people that are capable of using, uh, large language models in, in their domains and in their languages.
- 9:38
Yeah. And we have spec, uh, specifically increased the capabilities, uh, in, uh, agents and codings, and especially we enhanced the support, uh, for MCP, which is, uh, really pop, popular, uh, very recently.
- 9:52
And, and these are two examples to show that, uh, our models, how to use tools during thinkings. Who will find that, uh, it can think while it, uh, it is using a function calls to use the tools, and it gets the feedback from the environment and it keeps thinking.
- 10:12
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.
- 10:24
It is really good for, like, uh, inference time scaling. And this is another example for the model to organize the desktop. So, uh, if it has the access to the file system, it can do things like that.
- 10:37
It thinks, uh, which tools should they use, and then it use the tools and gets the feedback and continues thinking while, uh, it, well, finish the u-using the tools and it finish the task and tell you that, well, you...
- 10:50
we have organized the de-desktop quite well. And these are two very simple e-, uh, examples to show that, uh, we have provided better and better support for, uh, agent capabilities.
- 11:02
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.
- 11:12
Yeah. And, uh, these are three features of, um,
- 11:18
Qwen 3. Yeah. Uh, we have open-weighted, uh, a lot of sizes, um, including two MoE models. Uh, a small MoE model has a total number of, uh, thirty billion and only activates three billion, but, uh, another one is two hundred and thirty-five billion activates, uh, twenty, twenty bi-, uh, twenty-two billion parameter models.
- 11:41
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 four billion parameter model, models, you can deploy it, uh, on mobile devices and for thirty-two billion parameter models that, uh, these are models that, uh, you really prefer, and especially thirty-two billion parameter model.
- 12:01
It is strong. Uh, it is, shows competitiveness, and you can use it for doing, doing RL. You can deploy it, uh, uh, in, in your lo-local environment as well.
- 12:12
So we also, uh, open-weighted, um, uh, dense models as well. But we believe that in this year, maybe in the next years, the f-, uh, future trend is belonging to the MoE models.
- 12:23
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, uh, for the MoE models as well.
- 12:34
Yeah. Besides building large language models, we are also building multi-modal models. We have focused quite a lot on vision language models and, uh, I think, uh, many of you maybe are using Qwen2-VL and now using Qwen2.5-VL, uh, which was released, uh, this year in January.
- 12:54
Uh, it achieves a very competitive performance, uh, in, uh, vision language, uh, benchmarks, uh, like understanding benchmarks like MMMU and also math benchmark like MathVista and also general, uh, a lot of general VQA, VQA benchmarks.
- 13:09
So it achieved a very good performance in the, uh, benchmarks for, uh, vision language understanding. Um, we also explored the capabilities of thinking for vision language models, so we have, uh, built QVQ as well, and we find inverse time scaling with larger, uh, maximum thinking length, which is equivalent, uh, to the thinking budget that, uh, I talked
- 13:32
before. So if you have a larger thinking budget, it will achieve better and better performance, uh, in reasoning tasks especially, uh, uh, like mathematics. Uh, even for vision language models, it shows, uh, similar features as well.
- 13:48
Um, but for multi-modal models, we... what we really would like to do is to build an omni model which accepts multi modalities, uh, for the input and also it is capable of generating multiple modalities, uh, like text, vision and audio.
- 14:04
But this time, um, this is not a, a per- perfect state, but I think it's good. It's a relatively a small model, but we are really proud of this attempt.
- 14:14
It's a s- a seven billion large language model, it's based on it, but it is capable of accepting three modalities, including text, vision. Vision includes, uh, images and videos, and also, uh, it can accept, uh, au- audio, and this time it can generate text and audio.
- 14:32
Maybe in the future, our models m- might be capable of generating images, high quality images and videos. That would be a truly omni models. Now, for this omni model, it is, uh, it can be used in, uh, voice chat and video chat and text chat as well.
- 14:48
Yeah. Um, it achieved state-of-the-art performance, uh, in audio tasks, uh, for, for, for the same size models, like seven billion, uh, parameter models. Uh, but what surprised us a little bit is that, uh, it can even achieve better performance in vision language understanding tasks in comparison with Qwen 2.5-VL seven billion.
- 15:09
Which means that, um, we can achieve very good performance for an omni model in vision language tasks. But a, a little bit that we had not done well, but we believe we can done, uh, we, we can do well, is that, uh, we should recover the performance drop in language tasks, uh, especially for its intelligence, uh, especially
- 15:33
for its, uh, agent tasks. I, I think we can recover it, uh, uh, by improving our data quality, improving our training, training methods. But now, um, there is still some room for improving, uh, uh, the model capabilities in different domains and different tasks.
- 15:49
Uh, and this is the, uh, omni models. And no matter what models, we keep open sourcing. We love open sourcing because open sourcing really helps us quite a lot, um, by, uh, the developers can give us some feedbacks to help us, uh, improve our models.
- 16:07
The interaction with the cape, uh, open source community, uh, makes us happy and makes us, um, we have more... We are more encouraged to build more good models to, for, for all of you.
- 16:20
Yeah. We have a lot of titled models, uh, in the open, uh, source communities, including LLMs and also coders, and Qwen 2.5 coders, uh, is something that, uh, uh, many people are using for local development.
- 16:34
And something that I can tell you that we are now building Qwen3 coders. I think you guys know about it, yeah? Um, we have many model sizes because w- we, we just believe that for each size there might be a lot of users.
- 16:48
And actually, there are. There are a lot of users, no matter very, very small models like zero point six billion parameter models. And there are large models, previously seventy-two billion dense models and now two hundred and thirty-five billion, uh, MoE models.
- 17:02
There are a lot of users that are using it. And they need quantized models, so we just provide quantized models in different formats, including GGUF, GBQ, AWQ, and MOX for Apple as well.
- 17:14
We try to use Apache 2.0 for most models, so you can just use it freely. You can change the models freely, uh, in your business. You don't need to worry about too many things.
- 17:26
You don't need to ask us, uh, for the permission. You just directly use it. We hope that large language models and large multi-model models or foundation models can help you create good applications.
- 17:39
And this is what we like to do. And, uh, as we are becoming popular and popular during these two years, so for Qwen models, uh, it should be supported by a lot of, maybe most, uh, re- uh, relative third-party frameworks and API platforms as well.
- 17:57
Yeah. And we are also building products. You know, we are also building products for, for you to interact with our models. We are building agents as well. And in our Qwen chat, just as I, uh, mentioned before, we have some, uh, very good features.
- 18:12
This is something that I really like, which is called WebDev. Um, by entering WebDev, you just need to insert or input a very simple prompt. Uh, things like create a Twitter website, and you'll find that, well, it, it just generates a code, uh, and then, uh, some effects like artifacts.
- 18:33
So you, you have a website, uh- To, to see how, 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.
- 18:45
You can also create, uh, a website, uh, for your product. For example, this is, mm, a very simple prompt. Create a sunscreen product introduction website, and you will have a very good website.
- 18:58
You can even click on the buttons as well. And for making a card, uh, not, not just making a website, just by making a card, well, it, it generates a card.
- 19:10
I, I really like the card. We, we often use the card in our Twitter. So, uh, 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.
- 19:25
You just based on the information and build a card for, for us. Yeah. Um, this is something I really, really like. WebDev, uh, uh, makes me more creative and helps me quite a lot to show our things to all...
- 19:40
to, to people all over the world. Yeah. We also have things like deep research. Uh, by doing deep research, um, you just need to ask it something to write a report, uh, in what you are interested in.
- 19:53
Mm, yeah, m-maybe like healthcare industry, maybe like artificial intelligence. Just ask. Give it a prompt, and it will ask you what you are going to focus on. You can tell it, and you...
- 20:07
or, 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.
- 20:21
And keep searching, and finally gives a comprehensive report to all of it. You can download our report via a PDF. We are still improving, uh, its quality by, uh, doing reinforcement learning to build a fine-tuned model specifically for deep research.
- 20:41
And we believe that there's still much room, uh, in this field. It is really hard to do reinforcement learning, uh, for this type. But, um, once you have built a good model for, uh, for this product, it will be really productive for people, uh, in their working life.
- 20:57
Yeah. So in the future, we will do many things. Uh, there are still a lot of things for us to do to achieve AGI, to build a really good foundation model and foundation, you know, agent for all of you.
- 21:11
And the first thing is that, uh, something really different from what other people think. We still believe that there is still much room in pre-training. Um, I'm happy that you, uh, you have shown your preference, uh, in our, uh, pre-training models.
- 21:27
Uh, but we still find that there are still a lot of good data we did not put into it. Uh, there are still a lot of data we did not clean it quite well, and we find that we can use multi-model data and to make the models more, uh, capable, uh, in doing different tasks in different domains.
- 21:46
And we also have synthetic datas, and maybe m-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.
- 22:01
And so there is still much room, uh, in pre-training to build a very good basis for the, uh, chatbot or agents, and this is the first thing. And the second thing is that, uh, the scaling laws, uh, there are some changes in scaling law.
- 22:18
Previously, we are scaling, uh, in the model sizes, in the pre-training data, but now we need to scale the compute, uh, in reinforcement learning. And we have a focus on long-horizon reasoning, uh, with the environment feedback.
- 22:32
So, uh, if you train the model, which is capable of, of interacting with the environment, keeps thinking, this will be something really, really, uh, competitive. So it will get the feedback from the environment, keeps thinking, it will become smarter and smarter with inference time scaling.
- 22:49
Yeah. So you'll find that it, it will generate very long context.
- 22:55
And you have very long context for your input, uh, especially when you have memory. So you need to scale your context. Maybe finally we are m-moving towards infinite context.
- 23:07
But now we need to fix the problems of one million tokens, uh, quite well, and then we are marching to ten million tokens, and then infinite context. So we are scaling context.
- 23:18
Uh, we are going to scale the con- context at least one million tokens this year for most of our models. Yeah. We're also going to scale modalities. Um, may-maybe scaling your modalities doesn't increase your intelligence, but, uh, if you scale your modalities, you can make your models more capable and more productive, uh, especially with the vision language
- 23:40
understanding. You, you... if you have vision language understanding capability, you can make a, like, GUA agent. But before that, if you, uh, have no vision capability, you can... it is almost impossible for you to make a GUA agent and do things like, uh, com- uh, computer use.
- 23:57
And, um, maybe there is still much room, uh, in, uh, scaling the modalities in either inputs and outputs. Uh, we are going to unifying, uh, understanding and gener- generation.
- 24:10
Uh, for example, for, for the image understanding and image generation. At the same time, well, just like GPT-4o, they generate, uh, very interesting, uh, and high quality, uh, images.
- 24:21
That, that is something what we are going to do as well. Yeah. Uh, 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.
- 24:40
We are actually training the agents, uh, not only with scaling with pre-training, but also scaling with RL, especially with the environment. We are actually training the agents. So, uh, I think we can say that we are now staying in the, uh, era of agents.
- 24:58
Yeah, that's all. Uh, thank you very much, uh, for listening to my talk. And if you are interested, uh, in Qwen, uh, shoot me an email and talk to me in X.
- 25:10
Um, yeah. Thanks a lot. Yeah.