# The Future of Qwen: A Generalist Agent Model

Junyang Lin

AI Engineer World's Fair 2025

[Talk page](https://ai.engineer/talks/b0xlsQ_6wUQ-future-qwen-generalist-agent-model) · [Video](https://www.youtube.com/watch?v=b0xlsQ_6wUQ)

## 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.

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?

[0:23](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=23s)

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](https://chat.qwen.ai/) 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](https://qwenlm.github.io/) 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.

### From instruction following to reasoning

[1:35](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=95s)

[**Qwen2.5-Max**](https://qwenlm.github.io/blog/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.

![Grouped bars compare Qwen2.5-Max with four other models on Arena-Hard, MMLU-Pro, GPQA-Diamond, LiveCodeBench and LiveBench.](https://ai.engineer/api/talk-essay-media/2ae69371ba0cd8fa84b7eac31d4592fadce1c5014c867cde2fbc36b3b57fa363)

Qwen2.5-Max benchmark comparisons across five evaluations.

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**](https://qwenlm.github.io/blog/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.

### Capability at different compute scales

[3:07](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=187s)

[**Qwen3**](https://qwenlm.github.io/blog/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.

### Choosing when to think

[5:19](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=319s)

**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.

### How much thinking is enough?

[6:36](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=396s)

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.

![Four plots for AIME24, AIME25, LiveCodeBench and GPQA Diamond show blue thinking-mode curves increasing with thinking budget, above dashed non-thinking baselines.](https://ai.engineer/api/talk-essay-media/75e0316f73ba65cfa9817e9b42b79bd5b5147f397c1c7abc2b70f742b08f580a)

Qwen3 performance rises with larger thinking budgets.

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.

### From multilingual answers to tool interaction

[8:58](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=538s)

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:

1. Determine which filesystem tool is needed for the next step.
2. Invoke the tool and receive its result.
3. Continue reasoning with that environmental feedback.
4. 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.

### A lineup for local experimentation

[11:18](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=678s)

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.

### Reasoning over visual inputs

[12:34](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=754s)

The same progression extends to vision-language models. **Qwen2-VL** was followed by [**Qwen2.5-VL**](https://qwenlm.github.io/blog/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**](https://qwenlm.github.io/blog/qvq-72b-preview/) 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.

![Scatter plot shows MathVision accuracy of 43.5%, 45.6%, 46.7% and 48.1% at maximum thinking lengths of 4, 8, 16 and 24 thousand tokens.](https://ai.engineer/api/talk-essay-media/7033c0ff38dea91724465ef0392f0bc95c946955e42ff9aed97de5f7d42b1636)

QVQ MathVision accuracy increases with maximum thinking length.

### Combining modalities without losing capability

[13:48](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=828s)

An omni model broadens both what the model can perceive and what it can produce. [**Qwen2.5-Omni**](https://arxiv.org/html/2503.20215v1) 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.

### Making the models usable

[15:49](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=949s)

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.

### WebDev: a prompt becomes a shareable artifact

[17:57](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1077s)

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.

### Deep research: planning, searching and writing

[19:40](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1180s)

**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:

1. Make a research plan.
2. Search step by step.
3. Write parts of the report incrementally.
4. Continue searching as the report develops.
5. 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.

### Scaling the training and environment of an agent

[20:57](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1257s)

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.

![Future work slide lists room in pretraining, reinforcement learning with environment feedback for long-horizon reasoning, scaling context, and scaling input and output modalities.](https://ai.engineer/api/talk-essay-media/8326161a23161079f9c38cb720a3056c51bbd2c15dc1271f374f313c9182eb20)

Future work includes scaling input and output modalities.

**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.

## Resources

- [Qwen Chat](https://chat.qwen.ai/) — The public Qwen chat interface, now presented as Qwen Studio.
- [Qwen3: Think Deeper, Act Faster](https://qwenlm.github.io/blog/qwen3/) — The original Qwen3 release overview, with model sizes, language coverage, hybrid thinking controls and deployment examples.
- [QwQ-32B and reinforcement learning](https://qwenlm.github.io/blog/qwq-32b/) — The Qwen team's account of outcome-based reinforcement learning for mathematical reasoning, coding and general capabilities.
- [QwQ-32B model card and weights](https://huggingface.co/Qwen/QwQ-32B) — Official weights and usage guidance for the 32-billion-parameter reasoning model.
- [Qwen2.5-Max release overview](https://qwenlm.github.io/blog/qwen2.5-max/) — Historical benchmark comparisons and training background for the large MoE model.
- [Qwen2.5-VL release overview](https://qwenlm.github.io/blog/qwen2.5-vl/) — Vision-language capabilities spanning documents, video, localization and computer use.
- [QVQ: To See the World with Wisdom](https://qwenlm.github.io/blog/qvq-72b-preview/) — The December 2024 introduction to QVQ's experimental visual reasoning.
- [Qwen2.5-Omni Technical Report](https://arxiv.org/html/2503.20215v1) — The original technical report, including task-specific comparisons with specialist vision and audio models.
- [Implementing Qwen3 thinking budgets](https://github.com/QwenLM/Qwen3/blob/main/docs/source/getting_started/thinking_budget.md) — Current example separating budget-limited reasoning from final-answer generation.
- [Qwen-Agent](https://github.com/QwenLM/Qwen-Agent) — Qwen's agent framework with tool calling, MCP integration and code-interpreter support.

## Slides

### Important links: Qwen Chat

Qwen Chat is the team's interface for interacting with its latest models. The speaker describes image and video uploads, plus voice and video chat.

[0:31](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=31s) [Important links slide showing the Qwen Chat link.](https://assets.ai.engineer/sites/aiecode2025/sha256/298a39c4fc9720dddbdfa8f38b344e89c9d79226a6caf249ecee136490d339fc.webp)

Important links

• Qwen Chat: https://chat.qwen.ai/

Interface text:
Junyang Lin

### Qwen Chat interface

Qwen Chat shows upload options for documents, images, video and audio, alongside voice and video chat and suggested tasks such as image generation, coding and summarization.

[0:41](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=41s) [Qwen Chat interface with upload options and suggested tasks.](https://assets.ai.engineer/sites/aiecode2025/sha256/e85c22f62bfa35b06459c73c49af5a32129249f29443e080ee9ffbefac867178.webp)

Good afternoon, QwQ
How can I help you today?

Thinking
Search
MCP

Upload document
Upload image
Upload video
Upload audio

Image Generation
Video Generation
Code
Make a plan
Analyze image
Summarize text
Get advice
Help me write
Brainstorm

Use voice and video chat

Interface text:
Qwen2.5-Max
Cancel default
Junyang Lin

### Important links: Chat and blog

Qwen Chat lets users interact with Qwen models; the blog provides technical details on new releases.

[0:59](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=59s) [Important links slide showing Qwen Chat and blog links over a faint blog page.](https://assets.ai.engineer/sites/aiecode2025/sha256/264c13470b6285a55f15573bff4bdc02f750673ac0741933c98d3522b6ca1081.webp)

Important links

• Qwen Chat: https://chat.qwen.ai/
• Blog: https://qwenlm.github.io/

QVQ-Max: Think with Evidence
Qwen2.5 Omni: See, Hear, Talk, Write, Do It All!
Qwen2.5-VL-32B: Smarter and Lighter

Interface text:
Junyang Lin

### Qwen blog

Qwen's blog provides technical details alongside new releases. The page lists posts on QVQ-Max, Qwen2.5 Omni, and Qwen2.5-VL-32B.

[1:06](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=66s) [Qwen blog page listing articles about QVQ-Max, Qwen2.5 Omni, and Qwen2.5-VL-32B.](https://assets.ai.engineer/sites/aiecode2025/sha256/a74d1ecab9fe29c12c595f622fc2aef6215affe6f4c038306e3884215ed9181d.webp)

Blog | 简体中文 |

QVQ-Max: Think with Evidence
QWEN CHAT GITHUB HUGGING FACE MODELSCOPE DISCORD Introduction Last December, we launched QVQ-72B-Preview as an exploratory model, but it had many issues. Today, we are officially releasing the first version of QVQ-Max, our visual reasoning model. Th...
March 28, 2025 · 4 min · 829 words · Qwen Team

Qwen2.5 Omni: See, Hear, Talk, Write, Do It All!
QWEN CHAT HUGGING FACE MODELSCOPE DASHSCOPE GITHUB PAPER DEMO DISCORD We release Qwen2.5-Omni, the new flagship end-to-end multimodal model in the Qwen series. Designed for comprehensive multimodal perception, it seamlessly processes diverse...
March 27, 2025 · 3 min · 496 words · Qwen Team

Qwen2.5-VL-32B: Smarter and Lighter
QWEN CHAT GITHUB HUGGING FACE MODELSCOPE DISCORD Introduction At the end of January this year, we launched the Qwen2.5-VL series of models, which received widespread attention and positive feedback from the community. Building on the Qwen2.5-VL series, w...
March 24, 2025 · 10 min · 2026 words · Qwen Team

Interface text:
Blog
Publication
Junyang Lin

### Qwen model collections

Qwen’s organization page groups its models and demos by capability, including Omni, vision-language, long-context and visual reasoning. The speaker says its Hugging Face checkpoints can be downloaded to try the models.

[1:25](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=85s) [Qwen organization page with model collections and team-member avatars.](https://assets.ai.engineer/sites/aiecode2025/sha256/6194e7688c354e7c2a257a9fcc55da081c7a37e4c3bf7f83dad8080e891b7977.webp)

Organization Card

Welcome to Qwen 👋

This is the organization of Qwen, which refers to the large language model family built by Alibaba Cloud. In this organization, we continuously release large language models (LLM), large multimodal models (LMM), and other AGI-related projects. Feel free to visit Qwen Chat and enjoy our latest models!

Collections 14

Qwen2.5-Omni
End-to-End Omni (text, audio, image, video, and natural speech interaction) ...
Qwen/Qwen2.5-Omni-7B
Any-to-Any · Updated 7 days ago · 171k · 1.46k
Qwen2.5 Omni 7B Demo 🏆
Generate text and speech responses from text, images, or audio input · 260
Qwen2.5-Omni Technical Report
Paper · 2503.20215 · Published 27 days ago · 139

Qwen2.5-VL
Vision-language model series based on Qwen2.5
Qwen2.5 VL 32B Instruct Demo
Chat with images and videos using Qwen · 86
Qwen2.5 VL 72B Instruct
Interact with Qwen2.5-VL-Chat model using text and files · 224
Qwen2.5-VL Technical Report
Paper · 2502.13923 · Published Feb 20 · 182

Qwen2.5-1M
The long-context version of Qwen2.5, supporting 1M-token context lengths
Qwen/Qwen2.5-14B-Instruct-1M
Text Generation · Updated Jan 29 · 40.2k · 305
Qwen/Qwen2.5-7B-Instruct-1M
Text Generation · Updated Jan 29 · 2.57M · 314
Qwen2.5-1M Technical Report
Paper · 2501.15383 · Published Jan 26 · 71

QVQ
QVQ: Qwen models for visual reasoning
Qwen/QVQ-72B-Preview
Image-Text-to-Text · Updated Jan 12 · 50.1k · 586
QVQ 72B Preview
Upload images and ask questions to get answers · 566

Interface text:
AI & ML interests
None defined yet.
Recent Activity
xiongwang new activity · 7 days ago
Qwen/Qwen2.5-Omni-7B: The PR referenced in documentati...
xiongwang new activity · 7 days ago
Qwen/Qwen2.5-Omni-7B: What is the prompt for ASR
xiongwang updated a model · 7 days ago
Qwen/Qwen2.5-Omni-7B
View all activity
Team members 137
+39
Collapse
Junyang Lin

### Qwen2.5-Max benchmarks

Qwen2.5-Max leads this five-model comparison on Arena-Hard and LiveBench; Claude-3.5-Sonnet-1022 scores highest on the other three benchmarks.

[1:34](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=94s) [Bar chart comparing Qwen2.5-Max with four models across five benchmarks.](https://assets.ai.engineer/sites/aiecode2025/sha256/1b702d89841290362113091552b5d72e1a11ecb6b77b7a997e3eb6db84368b38.webp)

Qwen2.5-Max

Legend: Qwen2.5-Max; DeepSeek-V3; Llama-3.1-405B-Inst; GPT-4o-0806; Claude-3.5-Sonnet-1022
Y-axis ticks: 0, 20, 40, 60, 80, 100

Values in legend order:
Arena-Hard: 89.4, 85.5, 69.3, 77.9, 85.2
MMLU-Pro: 76.1, 75.9, 73.3, 77.0, 78.0
GPQA-Diamond: 60.1, 59.1, 51.1, 53.6, 65.0
LiveCodeBench (24.08-24.11): 38.7, 37.6, 30.2, 35.1, 38.9
LiveBench (2024-08-31): 62.2, 60.5, 53.2, 56.0, 60.3

Interface text:
Junyang Lin

### Qwen2.5-Max leaderboard

The leaderboard highlights Qwen2.5-Max at Rank* (UB) 12, with an Arena Score of 1340 and a Style Control rank of 14.

[2:05](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=125s) [Leaderboard with the Qwen2.5-Max row highlighted at rank 12.](https://assets.ai.engineer/sites/aiecode2025/sha256/eec1f07422d8f95fe37f8969851911b6a2208c4104b806fe8b67625ab2f791df.webp)

Qwen2.5-Max

Category: Overall
Apply filter: Style Control; Show Deprecated
Overall Questions
#models: 225 (100%)  #votes: 2,871,517 (100%)

Rank* (UB) | Rank (StyleCtrl) | Model | Arena Score | 95% CI | Votes | Organization | License
1 | 1 | Gemini-2.5-Pro-Exp-03-25 | 1439 | +7/-5 | 9013 | Google | Proprietary
2 | 2 | ChatGPT-4o-latest_(2025-03-26) | 1407 | +6/-6 | 8261 | OpenAI | Proprietary
2 | 4 | Grok-3-Preview-02-24 | 1402 | +5/-3 | 14849 | xAI | Proprietary
2 | 2 | GPT-4.5-Preview | 1398 | +5/-6 | 14520 | OpenAI | Proprietary
2 | 4 | Gemini-2.5-Flash-Preview-04-17 | 1392 | +10/-13 | 3325 | Google | Proprietary
5 | 9 | Gemini-2.0-Flash-Thinking-Exp-01-21 | 1380 | +4/-3 | 26309 | Google | Proprietary
5 | 5 | Gemini-2.0-Pro-Exp-02-05 | 1380 | +5/-4 | 20127 | Google | Proprietary
5 | 4 | DeepSeek-V3-0324 | 1372 | +9/-6 | 5888 | DeepSeek | MIT
9 | 6 | DeepSeek-R1 | 1358 | +4/-5 | 16077 | DeepSeek | MIT
9 | 14 | Gemini-2.0-Flash-001 | 1354 | +3/-4 | 22201 | Google | Proprietary
10 | 6 | o1-2024-12-17 | 1350 | +3/-3 | 28990 | OpenAI | Proprietary
11 | 14 | Gemma-3-27B-it | 1342 | +5/-6 | 9976 | Google | Gemma
12 | 14 | Qwen2.5-Max | 1340 | +4/-4 | 21169 | Alibaba | Proprietary
12 | 11 | o1-preview | 1335 | +3/-3 | 33170 | OpenAI | Proprietary
15 | 14 | o3-mini-high | 1325 | +4/-4 | 17988 | OpenAI | Proprietary
15 | 17 | DeepSeek-V3 | 1318 | +4/-3 | 22846 | DeepSeek | DeepSeek
15 | 21 | QwQ-32B | 1316 | +6/-6 | 7735 | Alibaba | Apache 2.0
16 | 20 | Gemini-2.0-Flash-Lite | 1311 | +5/-5 | 22122 | Google | Proprietary

Interface text: Junyang Lin

### QwQ-32B benchmarks

QwQ-32B scores 79.5 on AIME24, close to DeepSeek-R1-671B’s 79.8, and leads the five models on LiveBench and BFCL. Lin attributes gains in reasoning tasks to reinforcement learning.

[2:18](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=138s) [Bar chart comparing QwQ-32B with four reasoning models across five benchmarks.](https://assets.ai.engineer/sites/aiecode2025/sha256/3fda9b76aa3872b292fa471634031dbf0afee63dc882f75fd1e4f3c20f093bb6.webp)

QwQ-32B

Legend: QwQ-32B; DeepSeek-R1-671B; OpenAI-o1-mini; DeepSeek-R1-Distill-Llama-70B; DeepSeek-R1-Distill-Qwen-32B
Y-axis ticks: 0, 20, 40, 60, 80, 100

Values in legend order:
AIME24: 79.5, 79.8, 63.6, 70.0, 72.6
LiveCodeBench (24.08-25.02): 63.4, 65.9, 53.8, 57.5, 57.2
LiveBench: 73.1, 71.6, 59.1, 57.9, 54.6
IFEval: 83.9, 83.3, 84.8, 79.3, 72.5
BFCL: 66.4, 60.3, 62.8, 49.3, 53.5

Interface text:
Junyang Lin

### QwQ-32B leaderboard

QwQ-32B appears at rank 15 on the leaderboard with an Arena score of 1,316. Its style-controlled rank is 21, and its listed license is Apache 2.0.

[2:55](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=175s) [Leaderboard with the QwQ-32B row highlighted at rank 15.](https://assets.ai.engineer/sites/aiecode2025/sha256/8bddb1cde4e87b058fec7e46b02aa91b356a913d7d93ec858103f240c51a619c.webp)

QwQ-32B

Category: Overall
Apply filter: Style Control; Show Deprecated
Overall Questions
#models: 225 (100%)  #votes: 2,871,517 (100%)

Rank* (UB) | Rank (StyleCtrl) | Model | Arena Score | 95% CI | Votes | Organization | License
1 | 1 | Gemini-2.5-Pro-Exp-03-25 | 1439 | +7/-5 | 9013 | Google | Proprietary
2 | 2 | ChatGPT-4o-latest (2025-03-26) | 1407 | +6/-6 | 8261 | OpenAI | Proprietary
2 | 4 | Grok-3-Preview-02-24 | 1402 | +5/-3 | 14849 | xAI | Proprietary
2 | 2 | GPT-4.5-Preview | 1398 | +5/-6 | 14520 | OpenAI | Proprietary
2 | 4 | Gemini-2.5-Flash-Preview-04-17 | 1392 | +10/-13 | 3325 | Google | Proprietary
5 | 9 | Gemini-2.0-Flash-Thinking-Exp-01-21 | 1380 | +4/-3 | 26309 | Google | Proprietary
5 | 5 | Gemini-2.0-Pro-Exp-02-05 | 1380 | +5/-4 | 20127 | Google | Proprietary
5 | 4 | DeepSeek-V3-0324 | 1372 | +9/-6 | 5888 | DeepSeek | MIT
9 | 6 | DeepSeek-R1 | 1358 | +4/-5 | 16077 | DeepSeek | MIT
9 | 14 | Gemini-2.0-Flash-001 | 1354 | +3/-4 | 22201 | Google | Proprietary
10 | 6 | o1-2024-12-17 | 1350 | +3/-3 | 28990 | OpenAI | Proprietary
11 | 14 | Gemma-3-27B-it | 1342 | +5/-6 | 9976 | Google | Gemma
12 | 14 | Qwen2.5-Max | 1340 | +4/-4 | 21169 | Alibaba | Proprietary
12 | 11 | o1-preview | 1335 | +3/-3 | 33170 | OpenAI | Proprietary
15 | 14 | o3-mini-high | 1325 | +4/-4 | 17988 | OpenAI | Proprietary
15 | 17 | DeepSeek-V3 | 1318 | +4/-3 | 22846 | DeepSeek | DeepSeek
15 | 21 | QwQ-32B | 1316 | +6/-6 | 7735 | Alibaba | Apache 2.0
16 | 20 | Gemini-2.0-Flash-Lite | 1311 | +5/-5 | 22122 | Google | Proprietary

Interface text:
Junyang Lin

### Qwen3

Lin introduces Qwen3, a newly released family of dense and MoE language models in multiple sizes.

[3:19](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=199s) [Qwen3 title on a purple background with a bear mascot.](https://assets.ai.engineer/sites/aiecode2025/sha256/e82de79e5af18ae9733458c6ca3964ad89ac8a4eeb718455f7294bb4c228775a.webp)

Qwen3
Qwen

Interface text:
Junyang Lin

### Qwen3 flagship benchmarks

Qwen3-235B-A22B activates 22B of its 235B parameters. It scores above o3-mini on all nine listed benchmarks; Gemini 2.5 Pro leads on six.

[3:31](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=211s) [Benchmark table comparing Qwen3-235B-A22B and Qwen3-32B with other models.](https://assets.ai.engineer/sites/aiecode2025/sha256/c49ad282c7c95d6dd037b02b72a72347000d2f3adc884cd0ef84a6b4a5df5bb7.webp)

Qwen3

Columns: Qwen3-235B-A22B (MoE) | Qwen3-32B (Dense) | OpenAI-o1 (2024-12-17) | Deepseek-R1 | Grok 3 Beta (Think) | Gemini2.5-Pro | OpenAI-o3-mini (Medium)

ArenaHard: 95.6 | 93.8 | 92.1 | 93.2 | - | 96.4 | 89.0
AIME’24: 85.7 | 81.4 | 74.3 | 79.8 | 83.9 | 92.0 | 79.6
AIME’25: 81.5 | 72.9 | 79.2 | 70.0 | 77.3 | 86.7 | 74.8
LiveCodeBench (v5, 2024.10-2025.02): 70.7 | 65.7 | 63.9 | 64.3 | 70.6 | 70.4 | 66.3
CodeForces (Elo Rating): 2056 | 1977 | 1891 | 2029 | - | 2001 | 2036
Aider (Pass@2): 61.8 | 50.2 | 61.7 | 56.9 | 53.3 | 72.9 | 53.8
LiveBench (2024-11-25): 77.1 | 74.9 | 75.7 | 71.6 | - | 82.4 | 70.0
BFCL (v3): 70.8 | 70.3 | 67.8 | 56.9 | - | 62.9 | 64.6
MultiIF (8 Languages): 71.9 | 73.0 | 48.8 | 67.7 | - | 77.8 | 48.4

1. AIME 24/25: We sample 64 times for each query and report the average of the accuracy. AIME’25 consists of Part I and Part II, with a total of 30 questions.
2. Aider: We didn’t activate the think mode of Qwen3 to balance efficiency and effectiveness.
3. BFCL: The Qwen3 models are evaluated using the FC format, while the baseline models are assessed using the highest scores obtained from either the FC or prompt formats.

Interface text:
Junyang Lin

### Qwen3 smaller-model benchmarks

Qwen3-30B-A3B activates 3B of its 30B parameters and beats QwQ-32B on 7 of 9 listed benchmarks. Qwen3-4B exceeds Qwen2.5-72B-Instruct on both AIME tests, but trails on ArenaHard.

[4:07](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=247s) [Benchmark table comparing Qwen3-30B-A3B and Qwen3-4B with other models.](https://assets.ai.engineer/sites/aiecode2025/sha256/40d151cce106d2153780e41a790c8ff89fb1ca134d2d337280cfa18f7f1d9f68.webp)

Qwen3

Models: Qwen3-30B-A3B (MoE) | QwQ-32B | Qwen3-4B (Dense) | Qwen2.5-72B-Instruct | Gemma3-27B-IT | DeepSeek-V3 | GPT-4o (2024-11-20)

ArenaHard: 91.0 | 89.5 | 76.6 | 81.2 | 86.8 | 85.5 | 85.3
AIME’24: 80.4 | 79.5 | 73.8 | 18.9 | 32.6 | 39.2 | 11.1
AIME’25: 70.9 | 69.5 | 65.6 | 15.0 | 24.0 | 28.8 | 7.6
LiveCodeBench (v5, 2024.10-2025.02): 62.6 | 62.7 | 54.2 | 30.7 | 26.9 | 33.1 | 32.7
CodeForces (Elo Rating): 1974 | 1982 | 1671 | 859 | 1063 | 1134 | 864
GPQA: 65.8 | 65.6 | 55.9 | 49.0 | 42.4 | 59.1 | 46.0
LiveBench (2024-11-25): 74.3 | 72.0 | 63.6 | 51.4 | 49.2 | 60.5 | 52.2
BFCL (v3): 69.1 | 66.4 | 65.9 | 63.4 | 59.1 | 57.6 | 72.5
MultiIF (8 Languages): 72.2 | 68.3 | 66.3 | 65.3 | 69.8 | 55.6 | 65.6

1. AIME 24/25: We sample 64 times for each query and report the average of the accuracy. AIME’25 consists of Part I and Part II, with a total of 30 questions.
2. Aider: We didn’t activate the think mode of Qwen3 to balance efficiency and effectiveness.
3. BFCL: The Qwen3 models are evaluated using the FC format, while the baseline models are assessed using the highest scores obtained from either the FC or prompt formats.

Interface text:
Junyang Lin

### Qwen3 thinking mode

Qwen3 combines thinking and non-thinking in one model. In thinking mode, it explores possibilities before answering.

[5:19](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=319s) [Slide introducing hybrid thinking modes and step-by-step reasoning.](https://assets.ai.engineer/sites/aiecode2025/sha256/e67eef09bbf7e5cf77532119d9da0c57d6f54d4d632425d0721ae5cbe7bb42d9.webp)

Qwen3

• Hybrid thinking modes
  • Thinking mode: step-by-step thinking and reasoning

Interface text:
Junyang Lin

### Qwen3 hybrid thinking modes

Qwen3 combines step-by-step reasoning and near-instant responses in one model. Prompts or hyperparameters control its behavior, while a thinking budget sets the maximum thinking tokens.

[6:08](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=368s) [Slide describing thinking and non-thinking modes and dynamic thinking budgets.](https://assets.ai.engineer/sites/aiecode2025/sha256/fd2cdfb5cc26ef7efadd0bfbf24477424f69bc386c065ddb76899714237f2364.webp)

Qwen3

• Hybrid thinking modes
  • Thinking mode: step-by-step thinking and reasoning
  • Non-thinking mode: near-instant response
• Dynamic thinking budgets

Interface text:
Junyang Lin

### Thinking budget and benchmark performance

Across four benchmarks, thinking-mode scores rise as the budget increases from 1K to 32K tokens. The budget caps thinking tokens; reaching the cap can truncate the thinking process.

[6:44](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=404s) [Four line charts compare thinking-mode performance at increasing token budgets with non-thinking baselines.](https://assets.ai.engineer/sites/aiecode2025/sha256/f4c8d4b570ce11b3a22e93eebcac4c218b1a0e149577cf7edfcc1fffc6c5c8e5.webp)

Qwen3

AIME24
Legend: Non-thinking Mode; Thinking Mode
Y-axis: Pass@1
Y-axis ticks: 40, 50, 60, 70, 80
X-axis ticks: 1, 2, 4, 8, 16, 32

AIME25
Legend: Non-thinking Mode; Thinking Mode
Y-axis ticks: 30, 40, 50, 60, 70, 80
X-axis ticks: 1, 2, 4, 8, 16, 32

LiveCodeBench (v5)
Legend: Non-thinking Mode; Thinking Mode
Y-axis: Pass@1
Y-axis ticks: 35, 40, 45, 50, 55, 60, 65
X-axis: Thinking Budget (K tokens)
X-axis ticks: 1, 2, 4, 8, 16, 32

GPQA Diamond
Legend: Non-thinking Mode; Thinking Mode
Y-axis ticks: 64, 66, 68, 70, 72
X-axis: Thinking Budget (K tokens)
X-axis ticks: 1, 2, 4, 8, 16, 32

* The Pass@1 scores are averaged over 64 (AIME24 and AIME25) or 16 (LCB and GPQA) samplings.

Interface text:
Junyang Lin

### Qwen3 and MCP: GitHub Stars

The demo asks Qwen3-32B to extract the QwenLM repositories page and plot repository stars as a bar chart, with fetch and code interpreter tools enabled.

[9:37](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=577s) [Demo screen introducing a task to fetch GitHub star counts and plot a bar chart.](https://assets.ai.engineer/sites/aiecode2025/sha256/407cf555fb8744db123380e365ad5acda4ff200270dc211d2c80754738945a61.webp)

Qwen3+MCP
Fetch GitHub stars and plot a bar chart

Interface text:
Chatbot
Qwen3-32B
I am Qwen3-32B!
插件
time-get_current_time
time-convert_time
fetch-fetch
code_interpreter
https://github.com/orgs/QwenLM/repositories Extract markdown content of this page, then draw a bar chart to display the number of stars.
点击上传
提交
通过 API 使用
使用 Gradio 构建
Settings
Junyang Lin

[9:49](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=589s) [Qwen3 and MCP demonstration introducing a request to chart GitHub repository stars.](https://assets.ai.engineer/sites/aiecode2025/sha256/c04beb5800b08b78b9488d88fc2d18037af9b75b7d912d2294b5562760bf48bb.webp)

Qwen3+MCP
Fetch GitHub stars and plot a bar chart

Interface text:
Chatbot
Qwen3-32B
I am Qwen3-32B!
插件
time-get_current_time
time-convert_time
fetch-fetch
code_interpreter
https://github.com/orgs/QwenLM/repositories Extract markdown content of this page, then draw a bar chart to display the number of stars.
点击上传
提交
Junyang Lin

## Full transcript

[0:03](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=3s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=14s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=23s) a series of large language models and large multi-modal 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=35s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=45s) You can use our multi-modal 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 WebDev and, um, deep research.

[0:59](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=59s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=73s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=88s) Yeah. Um, so feel free to, uh, check our websites and enjoy them. Yep.

[1:35](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=95s) Um, in this year, just before the Spring Festival, we have released a very good instruction-tuned models. I think, uh, it's a very good basis, uh, for large language models, which is Qwen2.5-Max.

[1:49](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=109s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=124s) Uh, but we, uh, we, uh, we believe that there are more potential for a large language model, not just becoming an instruction-tuned model, but it can, uh, be smarter and smarter with reinforcement learning.

[2:19](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=139s) 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 AIME twenty, uh, twenty-four.

[2:41](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=161s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=176s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=187s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=201s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=213s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=226s) 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 2.5 Pro.

[4:03](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=243s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=255s) Although it has thirty billion total parameter models, on-only activate three billion parameter models. But if you compare the performance with QwQ-32B, the MoE models, which only activates three billion parameter models, even can outcompete QwQ-32B in some tasks.

[4:35](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=275s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=289s) 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 Qwen2.5-72B.

[5:07](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=307s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=319s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=339s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=354s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=368s) Yeah. And for non-thinking mode, it is just the traditional, uh, instruction-tuned model. It's just like a chatbot without thinking, without lagging. It is just provide the answer. It is near instant.

[6:23](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=383s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=396s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=411s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=431s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=441s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=453s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=470s) And, uh, with a very small thinking budget, uh, for example, in AIME 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=484s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=500s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=511s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=521s) 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 Qwen3 supports over one hundred and n-nineteen languages and dialects.

[8:58](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=538s) In Qwen2.5, 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=554s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=565s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=578s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=592s) And, and these are two examples to show that, uh, our models, how to use tools during thinkings. You 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=612s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=624s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=637s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=650s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=662s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=672s) Yeah. And, uh, these are three features of, um,

[11:18](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=678s) Qwen3. 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=701s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=721s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=732s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=743s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=754s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=774s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=789s) 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 inference time scaling with larger, uh, maximum thinking length, which is equivalent, uh, to the thinking budget that, uh, I talked

[13:32](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=812s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=828s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=844s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=854s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=872s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=888s) 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 Qwen2.5-VL seven billion.

[15:09](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=909s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=933s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=949s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=967s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=980s) Yeah. We have a lot of titled models, uh, in the open, uh, source communities, including LLMs and also coders, and Qwen2.5-Coder, uh, is something that, uh, uh, many people are using for local development.

[16:34](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=994s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1008s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1022s) 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, GPTQ, AWQ, and MLX for Apple as well.

[17:14](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1034s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1046s) 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-modal models or foundation models can help you create good applications.

[17:39](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1059s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1077s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1092s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1113s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1125s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1138s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1150s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1165s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1180s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1193s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1207s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1221s) And keep searching, and finally gives a comprehensive report to all of you. 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1241s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1257s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1271s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1287s) 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-modal data and to make the models more, uh, capable, uh, in doing different tasks in different domains.

[21:46](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1306s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1321s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1338s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1352s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1369s) Yeah. So you'll find that it, it will generate very long context.

[22:55](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1375s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1387s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1398s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1420s) understanding. You, you... if you have vision language understanding capability, you can make a, like, GUI agent. But before that, if you, uh, have no vision capability, you can... it is almost impossible for you to make a GUI agent and do things like, uh, com- uh, computer use.

[23:57](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1437s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1450s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1461s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1480s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1498s) 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](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1510s) Um, yeah. Thanks a lot. Yeah.
