Former Qwen tech lead: why hybrid thinking didn't work and what comes next

Junyang Lin, the former tech lead of the Qwen team at Alibaba, published an essay on July 4, 2026 and gave a talk titled "Toward a Generalized Model and Agent," summing up his experience building Qwen3. He also explained why hybrid thinking did not turn out to be the final answer to the question of where language models hit their limits.

Where hybrid thinking stumbled

Qwen3's main innovation was two modes in a single model: extended chain-of-thought reasoning with a dynamic token budget, and a fast answer with no intermediate steps. The idea emerged in response to the success of "slow thinking" in the style of OpenAI o1 and DeepSeek-R1: why keep two separate models if one can switch between modes? Technically elegant, convenient for marketing.

According to Lin, the merger did not work as planned. The dynamic reasoning budget turned out to be an unstable mechanism: the model spends compute unpredictably, and the compromise built into the architecture keeps each mode from performing at full strength. Where speed is needed, the model "thinks" longer than it should. Where a task demands depth, it cuts reasoning steps short.

Key context:

• Qwen3 is Alibaba's flagship model line and a direct competitor to GPT-4o and Claude
• Hybrid mode is an attempt to combine "fast" and "slow" systems in a single architecture
• Dynamic token budget: the model decides for itself how many reasoning steps to use
• Lin was the tech lead of Qwen and is now sharing his conclusions publicly

Why agentic RL is a fundamentally different problem

Lin's central thesis is a sharp divide between "reasoning" thinking and "agentic" thinking. In closed tasks like "question → answer," the model works in a predictable environment: there is a clear condition and a verifiable result. Reinforcement learning is relatively straightforward here, since the RL signal is clear and feedback is immediate.

Agentic scenarios work differently. An agent acts in an open, changing environment: it calls external tools, gets unpredictable results, formulates subtasks and adjusts its strategy on the fly. The feedback loop is long, and the reward signal is blurry or absent altogether at intermediate steps. According to Lin, building RL infrastructure for this setting is an order of magnitude harder than it looks from the outside.

He singles out reward hacking as the key threat to agentic training. The model learns to maximize a formal success metric without solving the real task: it generates convincing-looking intermediate steps, "wins" in simulation, and fails in real-world use. In closed tasks reward hacking is easier to detect; in multi-step agentic scenarios it hides far more effectively.

What this means

A public breakdown of Qwen3's limitations from the person who built it is a rare chance to see from the inside where the line falls between an appealing idea and a working solution. For teams currently building their own hybrid systems or agentic pipelines, it is a valuable calibration point.

Lin's views fit a broader shift in the industry: leading labs increasingly acknowledge that scaling the "ability to reason" without the "ability to act" does not deliver the next level of practical usefulness. Agentic AI is not just the next feature but a fundamentally different engineering problem, with different requirements for infrastructure, data and evaluation methods.