The Agentic AI Revolution: Qwen’s Former Lead on Why Thinking Models Evolve into Action-Oriented Agents

A Paradigm Shift: From Models to Agents

For readers tracking the shift, In the rapidly evolving landscape of artificial intelligence, a significant shift in perspective is emerging, championed by none other than Junyang Lin, the former technical lead of Alibaba’s influential Qwen project. Having stepped down in March 2026 to pursue independent research, Lin has articulated a compelling new thesis: the AI community is moving beyond merely training sophisticated models to developing autonomous, action-oriented agents.

Meanwhile, Lin’s insights, drawn from his talk ‘Qwen: Towards a Generalist Model / Agent’ and a subsequent detailed blog post, challenge conventional ‘hybrid thinking’ approaches in large language models (LLMs). He argues that while hybrid models attempted to merge different modes of intelligence, the future lies in agents that think not just to reason, but to *act* in complex environments.

The Evolution of Qwen: A Journey Through Hybrid Models

Lin’s tenure at Qwen saw the development of a diverse family of models, showcasing impressive capabilities across various benchmarks. His talk provided a comprehensive tour, highlighting models like QwQ-32B, Qwen2.5-Max, Qwen3, Qwen2.5-VL, and Qwen2.5-Omni. These models consistently demonstrated strong performance against contemporaries such as DeepSeek-R1, Grok 3 Beta, Gemini 2.5 Pro, and OpenAI’s o-series.

In practical terms, A key focus during the Qwen3 development was the concept of ‘hybrid thinking modes.’ This involved a deliberate attempt to combine a step-by-step reasoning mode, ideal for complex problem-solving, with a non-thinking mode designed for near-instant, direct responses. Qwen3 also introduced dynamic thinking budgets, allowing users to control the computational resources allocated to a model’s reasoning process, alongside significantly expanded multilingual support to 119 languages and dialects.

Inside Qwen3’s Architectural Innovations

The Qwen3 family spanned a wide range of model sizes, from a compact 0.6 billion parameters up to a massive 235 billion. Architectural details revealed interesting design choices:

• Smaller dense models (e.g., Qwen3-0.6B, Qwen3-1.7B) often tied input and output embeddings and utilized a 32K context window.
• Larger dense and Mixture-of-Experts (MoE) models (e.g., Qwen3-8B, Qwen3-235B-A22B) typically dropped embedding tying and extended their context window to 128K, enabling deeper understanding and generation.
• The MoE variants, like Qwen3-30B-A3B and Qwen3-235B-A22B, were designed to activate a subset of their total experts (e.g., 8 out of 128) per token, optimizing for efficiency and specialized processing.

For example, These models were also made available in various quantized formats, including GGUF, GPTQ, AWQ, and MLX, all under the permissive Apache 2.0 license, promoting widespread adoption and experimentation.

The Challenge of Hybrid Thinking: Why Merging Intelligence is Complex

While hybrid thinking appeared promising as a feature, Lin’s later reflections illuminated the profound difficulties in its implementation. He notes a fundamental tension: a strong instruct model thrives on directness and low latency, whereas a strong thinking model requires ample tokens and time for deliberation. Attempting to merge these conflicting objectives carelessly often leads to degradation in both areas – reasoning becomes bloated, and direct instruction loses its crispness.

That said, Qwen3’s initial approach to merging these modes involved a sophisticated four-stage post-training pipeline, incorporating techniques like long-CoT cold starts and reasoning RL. However, by late 2025, the Qwen 2507 line pivoted to shipping separate Instruct and Thinking variants, suggesting the challenge was more rooted in data and alignment than in model architecture alone.

Lin views Anthropic’s approach with Claude 3.7 Sonnet (a hybrid with user-set thinking budgets) and Claude 4 (interleaving reasoning with tool use) as a valuable corrective, emphasizing that the length of a reasoning trace doesn’t equate to intelligence. True thinking, he argues, must be shaped by the target workload, not merely by benchmarks.

From Deliberation to Action: Embracing Agentic Thinking

Lin draws a clear line between two eras of AI development. The first, ‘reasoning thinking,’ exemplified by models like o1 and DeepSeek-R1, focused on internal deliberation. This era taught the field the importance of deterministic, verifiable rewards (especially in math, code, and logic) for reinforcement learning (RL) and transformed RL into a systems challenge of large-scale rollouts and verification.

Interestingly, The next era, in Lin’s compelling framework, is ‘agentic thinking.’ This paradigm shifts the focus from internal monologue to active interaction with the world. An agent thinks *in order to act*, formulating plans, deciding when to execute, utilizing tools, interpreting environmental feedback, and dynamically revising its strategies. It’s defined by a closed-loop interaction, where thinking is a means to achieve sustained progress in an environment.

Reasoning vs. Agentic: A Core Distinction

Lin highlights critical differences between these two modes:

• Judged By: Reasoning thinking is assessed by the quality of internal deliberation; agentic thinking by sustained progress while acting.
• Reward Signal: Reasoning relies on verifiable answers (math, code, logic); agentic thinking on task success in an interactive environment.
• Core Object of Training: Reasoning trains the model; agentic thinking trains the model *plus its environment* (the harness).
• Infra Bottleneck: Reasoning faces bottlenecks in rollouts, verification, and stable policy updates; agentic thinking struggles with tool servers, sandboxes, and train-serve decoupling.
• Main Failure Mode: Reasoning can lead to verbose, low-value reasoning traces; agentic thinking risks reward hacking through tool access and environment leaks.

Key Capabilities of Agentic Thinking:

• Deciding precisely when to stop thinking and take an action.
• Strategically choosing which tools to invoke and in what sequence.
• Incorporating noisy or incomplete observations from the environment.
• Adapting and revising plans after encountering failures.
• Maintaining coherence and continuity across numerous turns and tool calls.

Real-World Impact: Agentic AI in Action

However, This distinction between reasoning and agentic thinking has profound implications for how AI systems are built and deployed:

• Coding Agents: Instead of merely suggesting a code patch, an agentic system can run a test harness, analyze real error messages, iteratively revise the code, and re-run tests until the suite passes. Here, thinking supports codebase navigation, error recovery, and tool orchestration.
• Deep Research: A reasoning model might generate a lengthy answer from its training data. An agentic system, however, can break down complex questions into sub-queries, perform dynamic searches, filter out unreliable sources, and deliver grounded, cited information. Qwen’s own Deep Research demo exemplifies this capability.
• Multi-Agent Orchestration: Lin anticipates a growing importance for ‘harness engineering,’ where an orchestrator agent plans and routes tasks. Specialized sub-agents then execute narrower functions, helping to manage context and complexity.

Practical Implementation: Qwen3’s Thinking Toggle

Qwen3 provided a concrete example of hybrid thinking through its `enable_thinking` flag in the chat template. Setting `enable_thinking=True` activates a step-by-step thinking mode, with reasoning often wrapped in a `<think>…</think>` block.

Conversely, `enable_thinking=False` yields a near-instant, non-thinking response. Furthermore, Qwen3 allowed for per-turn control via soft switches like `/think` or `/no_think` appended to user messages, forming the basis for dynamic thinking budgets.

The Infrastructure Hurdle: Building for Agentic RL

Meanwhile, Lin underscores that the shift to agentic AI presents significant engineering challenges, particularly for reinforcement learning (RL) infrastructure. In reasoning RL, rollouts are often self-contained trajectories with relatively clean evaluators. Agentic RL, by contrast, requires the policy to operate within a complex ‘harness’ comprising tool servers, browsers, terminals, and sandboxes.

This intricate environment mandates a crucial requirement: the clean decoupling of training and inference. Without it, rollout throughput collapses. Imagine a coding agent stalling inference while waiting for a live test execution – this starves training resources and significantly underutilizes GPUs.

Lin argues that the focus of optimization must shift from data diversity (the SFT era) to *environment quality* for agentic systems. This includes ensuring stability, realism, comprehensive coverage, and robust resistance to exploitation.

In practical terms, The central risk in this new paradigm is ‘reward hacking.’ As models gain real-world tool access, the attack surface for spurious optimization expands dramatically, making it harder to ensure agents genuinely achieve desired outcomes rather than exploiting loopholes in their environment or reward structure.

Key Takeaways from Junyang Lin’s Vision

• Junyang Lin’s departure from Qwen in March 2026 marks his transition to an independent researcher, focused on the future of AI.
• His core thesis is a fundamental shift from merely training models to developing and training sophisticated AI agents.
• Agentic thinking is defined by sustained action and progress within an interactive environment, rather than solely by internal deliberation.
• Building for agentic RL demands decoupled train-serve infrastructure and high-quality, robust environments, moving beyond simple verifiable rewards.
• The expanded tool access for agents makes ‘reward hacking’ a paramount challenge and central risk in this evolving AI landscape.

Expert Perspective

From an industry angle, the clearest signal around Agentic AI is how it may influence thinking. The story reads less like a one-day spike and more like a marker of broader movement.

The next phase will depend on how quickly teams, regulators, or customers react. In practice, that gives Agentic AI room to reshape expectations across qwen3 over the near term.

For readers focused on practical impact, the best next step is to watch what changes around models once attention turns into execution.

Frequently Asked Questions

Why does Agentic AI matter right now?

A Paradigm Shift: From Models to AgentsFor readers tracking the shift, In the rapidly evolving landscape of artificial intelligence, a significant shift in perspective is emerging, championed by none other than Junyang Lin, the former technical lead of Alibaba’s influential Qwen project.

What broader change could Agentic AI signal?

Having stepped down in March 2026 to pursue independent research, Lin has articulated a compelling new thesis: the AI community is moving beyond merely training sophisticated models to developing autonomous, action-oriented agents.Meanwhile, Lin’s insights, drawn from his talk ‘Qwen: Towards a Generalist Model / Agent’ and a subsequent detailed blog post, challenge conventional ‘hybrid thinking’ approaches in large language models (LLMs).

What should the market watch next around Agentic AI?

He argues that while hybrid models attempted to merge different modes of intelligence, the future lies in agents that think not just to reason, but to *act* in complex environments.The Evolution of Qwen: A Journey Through Hybrid ModelsLin’s tenure at Qwen saw the development of a diverse family of models, showcasing impressive capabilities across various benchmarks.

Source: https://www.marktechpost.com/2026/07/04/qwens-former-lead-on-what-hybrid-thinking-got-wrong-and-why-he-now-backs-agents/