Junyang Lin, the former technical lead of Alibaba’s Qwen project, has published a detailed critique of the hybrid thinking approach that powered the Qwen3 generation of models — and laid out why he believes agentic thinking, not longer reasoning chains, is the correct path forward.
Lin stepped down from his role at Alibaba in March 2026 and now operates as an independent researcher. His analysis, delivered first in a talk titled Qwen: Towards a Generalist Model / Agent and expanded in a subsequent blog post, represents one of the most detailed insider perspectives yet on the architectural and training trade-offs behind a major model family.
The hybrid thinking problem
Qwen3 shipped with two distinct modes: a thinking mode that produces step-by-step reasoning chains, and a non-thinking (instruct) mode for near-instant responses. The idea was to give users the best of both worlds in a single model.
Lin argues this fusion is fundamentally harder than it looks. Instruct mode rewards directness, brevity, and low latency. Thinking mode rewards spending additional tokens on hard problems. Merging both into one model, he said, degrades both.
“Longer reasoning traces do not make a model smarter,” Lin wrote. “Thinking should be shaped by the target workload, not by the benchmark.”
Qwen3 used a four-stage post-training pipeline — long chain-of-thought cold start, reasoning reinforcement learning, and “thinking mode fusion” — to attempt the merge. But later releases shipped separate Instruct and Thinking variants, effectively conceding that the unified approach was not working as intended.
Lin framed the difficulty as a data problem rather than a model architecture problem. “This is a data problem more than a model problem,” he said, pointing to the difficulty of constructing training data that serves both modes well simultaneously.
He acknowledged Anthropic’s approach with Claude 3.7 Sonnet, which exposes a user-configurable thinking budget, as a “useful corrective” to the field’s rush toward hybrid models.
From reasoning to agentic thinking
Lin’s more significant argument concerns what comes next. He divides recent AI progress into two eras:
Era 1 — Reasoning thinking, exemplified by OpenAI’s o-series and DeepSeek-R1. Reinforcement learning requires deterministic, verifiable rewards, which naturally biases training toward mathematics, code, and formal logic. RL becomes a systems engineering problem — large-scale rollouts and verification pipelines.
Era 2 — Agentic thinking, which Lin argues is the current frontier. The goal shifts from producing a correct answer to sustaining progress while acting in an interactive environment. An agent must decide when to stop deliberating and take action, which tool to invoke and in what order, how to incorporate noisy feedback from the world, and how to revise plans after failures.
“Thinking to answer is not the same as thinking to act,” he wrote.
The comparison table from his blog post draws the distinction sharply:
| Dimension | Reasoning thinking | Agentic thinking |
|———–|——————-|——————|
| Judged by | Quality of internal deliberation before an answer | Whether progress is sustained while acting |
| Reward signal | Verifiable answers (math, code, logic) | Task success in an interactive environment |
| Core training object | The chain of thought itself | The policy that selects actions, reads environment feedback, and adapts |
Lin expects “harness engineering” — the orchestration layer that routes work between specialized sub-agents — to matter more than raw model capability in the coming years. A single generalist model may eventually be less important than a system that can plan, delegate, verify, and recover.
The Qwen family itself demonstrated this trajectory. The talk’s closing slide ends with a single line: “Training models -> training agents.”
Concrete hooks for developers
Lin’s analysis also offers tangible guidance for practitioners. Qwen3’s `enable_thinking` flag, which switches between modes in the chat template, is the kind of ergonomic detail he believes the field should focus on simplifying:
“`python
from transformers import AutoModelForCausalLM, AutoTokenizer
name = “Qwen/Qwen3-8B”
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype=”auto”, device_map=”auto”)
enable_thinking=True -> step-by-step reasoning mode
enable_thinking=False -> near-instant, non-thinking mode
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True,
enable_thinking=True)
“`
For agentic deployments, Lin recommends multi-agent orchestration: an orchestrator model plans and routes work, while specialized sub-agents execute narrower tasks. This architecture controls context pollution — a problem that worsens as agentic tasks span many turns and tool calls — better than monolithic agents.
Qwen’s own Deep Research demo, which decomposes research questions into sub-queries, calls external search, evaluates source quality, and returns grounded citations, is presented as an early example of the agentic paradigm in action.
The Qwen3 family, released under Apache 2.0, spans model sizes from 0.6 billion to 235 billion parameters, with MoE variants (30B-A3B and 235B-A22B) activating 8 of 128 experts per token. Quantized formats including GGUF, GPTQ, AWQ, and MLX are available.
Sources: Qwen’s Former Lead on What Hybrid Thinking Got Wrong and Why He Now Backs Agents (MarkTechPost, July 2026); Qwen: Towards a Generalist Model / Agent (talk) (YouTube); Alibaba’s Qwen tech lead steps down (TechCrunch, March 2026); From ‘Reasoning’ Thinking to ‘Agentic’ Thinking (Lin’s blog)