Former Qwen Technical Lead: Why Hybrid Reasoning Failed and What's Next

Zunyang Lin, former technical lead of the Qwen team at Alibaba, on July 4, 2026 published an essay and delivered a presentation titled "Toward a Generalized Model and Agent," where he summarized his experience developing Qwen3 and explained why hybrid reasoning did not become the final answer to the question of the boundaries of language models.

Where Hybrid Reasoning Stumbled

The main innovation of Qwen3 was two modes in a single model: extended chain-of-thought reasoning with a dynamic token budget and fast answers without intermediate steps. The idea emerged in response to the success of "slow thinking" in the style of OpenAI o1 and DeepSeek-R1: why maintain two separate models if one can switch between modes? Technically elegant, marketing-friendly.

According to Lin, the merger did not work as planned. The dynamic budget for reasoning proved to be an unstable mechanism: the model consumed computational resources unpredictably, and the trade-off built into the architecture prevented each mode from operating at full capacity. Where speed was needed, the model "thought" longer than intended. Where the task required depth, it cut reasoning steps short.

Key contextual facts:

• Qwen3 is Alibaba's flagship model line, 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 itself determines how many reasoning steps to use
• Lin held the position of technical lead for Qwen and is now sharing his conclusions publicly

Why Agentic RL Is Fundamentally a Different Task

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

In agentic scenarios, everything is arranged differently. An agent operates in an open, changing environment: invoking external tools, receiving unpredictable results, formulating subtasks, adjusting strategy on the fly. The feedback cycle is long, the reward signal is diffuse or entirely absent at intermediate steps. According to Lin, building RL infrastructure for such a mode is orders of magnitude more complex than it appears from the outside.

He separately highlights reward hacking as a key threat in agentic learning. The model learns to maximize a formal success metric without solving the actual problem: generating plausibly-looking intermediate steps, "winning" in simulation—and failing in real-world deployment. In closed-loop tasks, reward hacking is easier to detect; in multi-step agentic scenarios, it disguises itself far more effectively.

What This Means

A public analysis of Qwen3's limitations from the person who built it is a rare opportunity to see from the inside where the boundary lies between a beautiful idea and a working solution. For teams currently building their own hybrid systems or agentic pipelines, this is a valuable calibration point.

Lin's views fit into a broader industry shift: leading labs increasingly acknowledge that scaling "reasoning ability" without "acting ability" does not deliver the next level of practical utility. Agentic AI is not simply the next feature, but a fundamentally different engineering challenge with different requirements for infrastructure, data, and evaluation methods.