# The Future of Qwen: A Generalist Agent Model

Junyang Lin, Alibaba Qwen | AI Engineer World's Fair 2025 | 25:14

Source: https://www.youtube.com/watch?v=b0xlsQ_6wUQ
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/the-future-of-qwen-a-generalist-agent-model
Published: 2025-06-03
Tags: agents, multimodal, open-models, reinforcement-learning

## TL;DR
- Qwen3 combines thinking and non-thinking behavior in one model, with prompts or hyperparameters controlling the mode and thinking budget.
- Qwen is adding tool use, MCP support, multimodal inputs, and product features such as WebDev and Deep Research to move from chatbots toward productive agents.
- The next stage of Qwen development will scale pre-training data, reinforcement learning with environment feedback, context length, and modalities.

## Summary
Junyang Lin presents Qwen3, Qwen2.5-VL, and Qwen2.5-Omni, then describes Alibaba Qwen's plan to build generalist agents. Qwen3 combines a fast instruction-following mode with a reasoning mode in one model. Its dynamic thinking budget lets developers trade token use against task performance. Lin also discusses support for 119 languages and dialects, coding, MCP, tool use during reasoning, and desktop tasks. The Qwen2.5-Omni model accepts text, images, video, and audio, and generates text and audio, although Lin says its language and agent performance still needs improvement. Qwen's products include WebDev, which creates deployable websites from prompts, and Deep Research, which plans searches and produces downloadable reports. For future models, the team plans to improve data and training methods, scale reinforcement learning with environmental feedback, support at least one million tokens of context, and unify more input and output modalities.

## Key ideas
### Reinforcement learning improves Qwen's reasoning performance
[02:19](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=139s)
Lin says Qwen's research moved beyond instruction tuning into reinforcement learning. He reports consistent gains on reasoning tasks such as mathematics and coding. In one example, a 32-billion-parameter model improved from about 65 to 80 on an evaluation as reinforcement learning progressed. He also says the model reached a competitive position in the arena rankings for a sustained period, despite being smaller than some competing systems. This work became part of the effort behind Qwen3, which Qwen released in several dense and mixture-of-experts sizes.

### Qwen3 puts fast responses and reasoning in one model
[05:16](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=316s)
Qwen3's hybrid thinking mode combines two behaviors in a single model. In thinking mode, the model reflects, explores possibilities, and produces an answer after its reasoning process. In non-thinking mode, it behaves like a traditional instruction-tuned chatbot and responds with low latency. Developers can control the behavior with prompts or hyperparameters. Lin presents this as a practical choice: a task that needs a quick answer can use the non-thinking mode, while a harder task can spend tokens on reasoning without requiring a separate model.

### A dynamic thinking budget trades tokens for task performance
[06:30](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=390s)
Qwen3 lets users set a maximum number of thinking tokens. If a task finishes in 8,000 tokens under a 32,000-token budget, the model can answer then. If the budget is only 4,000 tokens and the task needs 8,000, the reasoning process is truncated. Lin says performance rises as the budget increases. On one evaluation, a small budget produced a result just above 40, while a 32,000-token budget produced more than 80. He gives 8,000 tokens as an example of a budget that could reach over 95 percent accuracy when that level is sufficient.

### Qwen3 is designed for smaller deployments and broad language coverage
[07:31](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=451s)
Qwen3 supports more than 119 languages and dialects, compared with 29 in Qwen2.5. Lin says this should make the models more useful for global applications, especially where existing open models have weak language coverage. The release includes dense and mixture-of-experts models across several sizes. A four-billion-parameter model has thinking capabilities and can be deployed on mobile devices. A 30-billion-parameter mixture-of-experts model activates three billion parameters, and Lin says it can outperform a 32-billion-parameter model on some tasks.

### Qwen3 can reason while using tools and interacting with an environment
[09:25](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=565s)
Qwen3 adds stronger support for agents, coding, and MCP. Lin shows examples in which the model thinks, calls a function, receives feedback from the environment, and continues thinking. In a desktop-organization example, the model chooses tools, uses the file system, checks the results, and continues until the task is complete. Lin connects this loop to inference-time scaling and says Qwen should become more productive than a simple chatbot. The model's ability to keep reasoning while acting is part of Qwen's move toward general-purpose agents.

### Qwen2.5-Omni handles several modalities but still has gaps
[12:52](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=772s)
Qwen2.5-Omni is a relatively small, seven-billion-parameter model that accepts text, images, video, and audio. It generates text and audio, and it can power voice chat, video chat, and text chat. Lin says it reaches state-of-the-art performance among models of the same size on audio tasks. It also performs better than Qwen2.5-VL 7B on some vision-language understanding tasks. He is candid about a weakness: the omni model loses some performance on language intelligence and agent tasks. Qwen plans to address this through better data quality and training methods.

### Qwen's open-source strategy includes many sizes and formats
[16:21](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=981s)
Qwen continues to open-source language models, multimodal models, coders, and quantized versions. Lin says users need both very small models and large models, so Qwen provides a range from 0.6 billion parameters through 235 billion-parameter models. Quantized releases include GGUF, GPTQ, AWQ, and MLX formats for Apple devices. Most models use the Apache 2.0 license, which Lin says lets businesses use and change them directly. He also expects more third-party framework and API support, and says Qwen is building Qwen3-Coder.

### The next phase scales agents through data, RL, context, and modalities
[20:32](https://www.youtube.com/watch?v=b0xlsQ_6wUQ&t=1232s)
Lin describes a shift from training models toward training agents. Qwen plans to improve pre-training with more data, cleaner data, multimodal data, synthetic data, and possibly reinforcement learning during pre-training. It also plans to scale compute for reinforcement learning, especially for long-horizon reasoning with environment feedback. Context length is another target: Lin says the team is working on one million tokens, then ten million and eventually infinite context, with at least one million tokens planned for most models that year. Qwen also plans to scale input and output modalities and unify understanding with generation.

## Notable quotes
- "We would like to combine all our efforts in our research and development to build a stronger next generation models." (03:06)
- "You can use thinking and non-thinking in a single model." (05:16)
- "The model is capable of thinking but it is also capable of interacting with the environment and it keeps thinking." (09:48)
- "We are moving from the era of training models to training agents." (24:44)

## Tools & references mentioned
- Qwen3
- Qwen2.5-VL
- Qwen2.5-Omni
- Qwen2.5-Max
- Qwen2.5-Coder
- QVQ
- DeepSeek-R1
- o1
- Gemini 2.5 Pro
- GPT-4
- MCP
- Qwen Chat
- WebDev
- Deep Research
- GitHub
- Hugging Face
- Apache 2.0
- GGUF
- GPTQ
- AWQ
- MLX

## Who should watch
- You are choosing between fast instruction-following and longer reasoning for an application, and want to see how Qwen3 exposes that tradeoff in one model.
- You are building a tool-using or coding agent and need a model that can reason through function calls, MCP, file-system access, and environmental feedback.
- You need an open model for a particular language, device size, quantization format, or multimodal input.

## Related talks

- [Scaling the Next Paradigm of Heterogeneous Intelligence](https://aietalks.com/talks/scaling-the-next-paradigm-of-heterogeneous-intelligence) (Adrian Bertagnoli, Callosum, 15:13)
- [MCP Agent Fine-tuning Workshop](https://aietalks.com/talks/mcp-agent-fine-tuning-workshop) (Ronan McGovern, Trellis Research, 35:30)
- [The Future of MCP](https://aietalks.com/talks/the-future-of-mcp) (David Soria Parra, Anthropic, 18:46)
- [Building Applications with AI Agents](https://aietalks.com/talks/building-applications-with-ai-agents) (Michael Albada, Microsoft, 15:50)
- [AIE Europe Keynotes & Coding Agents](https://aietalks.com/talks/aie-europe-keynotes-coding-agents) (Tejas Kumar, AI Engineer & Omar Sanseviero, Google DeepMind & David Soria Parra, Anthropic & Ido Salomon, MCP Apps & Mario Zechner, Pi & Armin Ronacher & Cristina Poncela Cubeiro, Earendil & Benjamin Dunphy, AI Engineer & David Gomes, Cursor & Matthias Luebken, TAVON & Sarah Chieng, Cerebras & Lawrence Jones, Incident.io & Luke Alvoeiro, Factory & Ben Burtenshaw, Hugging Face & Michael Richman, Cmd+Ctrl & Liam Hampton, Microsoft & Tuomas Artman, Linear & Gergely Orosz, The Pragmatic Engineer & Jacob Lauritzen, Legora & Peter Gostev, Arena AI & swyx, AI Engineer, 9:09:51)