Junyang Lin - Leader of the Qwen LLM team, Alibaba Cloud. Tracked across 66 mentions in podcasts and expert conversations analyzed by Sonic.

Junyang Lin

AI researcher and leader of the Alibaba Qwen large language model team.

Claims

As speaker

Mentioned in

Key positions and views

The next frontier in AI is 'agentic thinking,' where models act and interact with environments, moving beyond the current focus on passive reasoning.

Competitive advantage in the agentic era will be determined by the quality of training environments, robust evaluation, and tight train-serve integration, not just model size.

Mixture-of-Experts (MoE) architectures are the future trend for large models, as they are critical for balancing performance with the inference and training efficiency required for real-world deployment.

The most significant technical challenges ahead are systems-level problems, particularly environment design, evaluator robustness, and preventing reward hacking in tool-using agents.

Scaling the volume of training data is more important for improving model capabilities than scaling the number of model parameters.

Timeline

Pre-Reasoning Era

Lin describes an industry focus on scaling pre-training, exemplified by the Qwen team's progression of training data from 2 trillion to 36 trillion tokens.

Emergence of Reasoning Models

Lin notes a significant transition in AI from scaling pre-training to scaling post-training. He cites OpenAI's o1 and DeepSeek-R1 as proof that 'thinking' could be an explicitly trained, first-class capability.

Qwen-3 Release

Alibaba releases the Qwen-3 series. Lin presents this as a notable attempt to create a unified model with a 'hybrid thinking mode' that combines chain-of-thought style reasoning and direct answering.

The Agentic Thesis

Lin articulates his core thesis that the industry is now moving beyond reasoning to 'agentic thinking,' where models act in environments. This marks a shift from training models to training agents.

Future Outlook

Lin outlines the next major research bottlenecks as environment design, evaluator robustness, and anti-cheating protocols. He details the Qwen roadmap, focusing on scaling context length to 1M+ tokens and building more complex agentic systems.

Podcast consensus on Lin

Points of consensus

▶The AI industry is undergoing a major paradigm shift, moving from an era focused on training models to one focused on training agents that interact with environments.Jul 2026

▶Mixture-of-Experts (MoE) architectures are a critical and growing trend for balancing state-of-the-art performance with the practical inference and training efficiency required for commercial services.Jul 2026

▶The focus of AI scaling laws is evolving from simply increasing model parameters and pre-training data to scaling the compute used for post-training, particularly reinforcement learning.Jul 2026

▶Developing advanced AI capabilities is as much an infrastructure and systems engineering challenge as it is a modeling challenge, requiring tighter integration between training and serving systems.Jul 2026

Points of debate

▶There is a tension between the goal of creating a single, unified model with both 'thinking' and 'instruct' modes (as in Qwen-3's hybrid mode) and the stated fundamental difficulty that their underlying data distributions and objectives are substantially different.Jul 2026

▶Lin's team is pursuing models with explicit 'thinking' capabilities, similar to OpenAI's o1, while he acknowledges that other major labs like Anthropic believe reasoning should be an integrated capability, not a separate mode.Jul 2026

▶A strategic conflict exists between the open-sourcing of smaller Qwen models to foster community and business opportunities, and keeping the largest, most powerful models like the 1T+ parameter Qwen 3 Max proprietary.Jul 2026

▶While Qwen's omni-models achieve state-of-the-art performance in specific modalities like audio, there is an acknowledged performance drop in core language and agentic tasks compared to specialized language-only models, highlighting a trade-off in the generalist vs. specialist approach.

Key themes

▶The Agentic Shift: From Reasoning to ActionJul 2026

Lin argues the AI industry is undergoing a fundamental transition from training models for reasoning to training agents for action. This 'agentic thinking' involves interacting with environments and updating plans based on feedback, a shift that redefines research bottlenecks and competitive advantages away from pure model scaling.

Investors should look for companies specializing in high-fidelity training environments and robust evaluation systems, as these will become critical, and potentially defensible, infrastructure in the next AI wave.

▶The New Scaling Laws: Post-Training and Data DominanceJul 2026

Lin repositions scaling laws, emphasizing that the focus is moving from pre-training data and model size to the compute used for reinforcement learning. He consistently highlights that scaling training data (from 2T to 36T tokens for Qwen) is more impactful than just increasing model parameters.

The value of proprietary, high-quality data and efficient RL infrastructure will likely surpass the value of raw model size, creating new opportunities for data-centric AI companies and those with strong systems engineering.

▶Architectural Pragmatism: The Rise of MoE and Hybrid Systems

Lin details the Qwen team's adoption of Mixture-of-Experts (MoE) and hybrid attention mechanisms as a pragmatic solution to balance performance with inference efficiency. This is driven by the real-world business need to serve models cost-effectively on Alibaba Cloud while continuing to scale capabilities.

The trend towards MoE and other efficiency-focused architectures suggests that the market for AI is maturing, with cost-of-service becoming as important as raw performance, favoring companies with strong systems engineering.

▶The Systems Challenge of Agentic AIJul 2026

Lin identifies the primary challenges of the agentic era as systems-level problems, not just algorithmic ones. These include preventing reward hacking, designing robust environments, creating cheat-proof evaluators, and cleanly decoupling training and inference systems to maintain throughput.

The complexity of building and deploying agentic AI creates a moat for vertically integrated players who can control the entire stack, from training environments to inference serving and feedback loops.

Source episodes

Suggested prompts

How does Junyang Lin's vision of an 'agentic era' dependent on training environments challenge the current paradigm focused on static datasets and pre-training? &nearr;Lin claims merging 'thinking' and 'instruct' modes is fundamentally difficult, yet Qwen-3 implements a 'hybrid thinking mode.' What are the specific trade-offs and architectural compromises involved in this approach? &nearr;Lin identifies reward hacking as the 'hardest challenge' for agentic AI with tool access. What specific anti-cheating protocols and evaluator robustness techniques does he allude to, and how might they be implemented? &nearr;Given Lin's emphasis on MoE for efficiency driven by Alibaba Cloud's business needs, how might the development trajectory of open-source models diverge from that of models developed for large-scale commercial API services? &nearr;

Key concepts

Agentic Thinking 3 ep Qwen Models 2 ep Mixture-of-Experts (MoE) 2 ep Scaling Laws 2 ep Context Length 2 ep Training Environments 2 ep Reasoning Models 1 ep Reinforcement Learning (RL) 1 ep Reward Hacking 1 ep

Notable quotes

“I believe the answer is agentic thinking: thinking in order to act, while interacting with an environment, and continuously updating plans based on feedback from the world.”

Junyang Lin · From Reasoning Thinking to Agentic Thinking

“The hardest challenge in training such systems is reward hacking. As soon as the model gets meaningful tool access, reward hacking becomes much more dangerous.”

Junyang Lin · From Reasoning Thinking to Agentic Thinking

“Previously, we're scaling in the model sizes and the pre-training data, but now we need to scale the compute in reinforcement learning.”

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

“Model size should be growing to be trillion parameter level. Now it's over 1 trillion parameters. Maybe later we will be 3 trillion, 5 trillion, maybe over 10 trillion.”

Junyang Lin · Junyang Lin Scaling Model Size and Context Length Towards Intelligence

Report last updated: Jul 28, 2026

Create a free account to see Junyang Lin's full intelligence report - every claim, the relationship network, and AI Q&A across all sources. No card needed.

Get started free

Back to Entities Intelligence Report