Junyang Lin stepped down as Alibaba’s Qwen technical lead on March 3, 2026. He now calls himself an independent researcher. His departure marks a shift away from hybrid thinking toward AI agents.
Hybrid thinking aimed to merge step-by-step reasoning with instant responses in one model. Qwen3 tried this with a complex four-stage post-training pipeline. But by late 2025, Alibaba split the models into separate Instruct and Thinking variants. Lin now agrees this was necessary.
Lin contrasts reasoning thinking—focused on internal deliberation—with agentic thinking, which plans, acts, and adapts in an environment. Agentic thinking demands more: deciding when to stop, choosing tools, revising plans, and staying coherent.
He argues the core training target differs. Reasoning thinking trains the model alone. Agentic thinking trains the model plus environment, requiring decoupled training and inference to avoid system stalls. Lin bluntly states, “Training models -> training agents.”
Alibaba’s Qwen family spans from 0.6 billion to 235 billion parameters. Qwen3 expanded multilingual support from 29 to 119 languages and dialects. It supports quantized formats like GGUF, GPTQ, AWQ, and MLX under Apache 2.0 licensing.
Qwen3.5, announced for Chinese New Year 2026, pushes parameter counts to 397 billion and supports 201 languages and dialects. It offers both open-weight and hosted versions, showing benchmark performance comparable to OpenAI, Anthropic, and Google DeepMind.
The April 2026 Qwen 3.6 27B model outperforms its 397B MoE predecessor on agentic coding tasks. It scores 59.3 on Terminal-Bench 2.0, matching Claude 4.5 Opus. It also beats the larger model on SWE-bench (77.2 vs. 76.2) and MMLU-Pro (86.2 vs. 84.8).
Qwen 3.6 uses a Gated DeltaNet architecture with linear attention to handle long contexts. It introduces native Multi-Token Prediction (MTP), doubling throughput from about 74 to over 140 tokens per second. Stability remains a problem; the Q4 MTP version fails 15% of the time. Q5 quantization improves accuracy.
The ‘preserve_thinking’ flag lets Qwen models remember previous deliberations across turns. This avoids “reasoning fresh every turn” and supports multi-turn agent loops. As the Tosea.ai team puts it, the model builds on earlier work instead of starting over.
Junyang Lin sees the future shifting sharply toward training agents able to act and adapt in real environments. He highlights the importance of resisting reward hacking and building high-quality environments for agentic reinforcement learning.
This evolution echoes Anthropic’s opposite hybrid approach, which Lin calls a useful corrective. Claude 3.7 Sonnet introduced user-set thinking budgets, while Claude 4 allows reasoning to interleave with tool use for coding and long tasks.
Running Frontier-tier AI models locally is no longer a fantasy. By mid-2026, a dual-3090 GPU setup suffices. The AI arms race now favors flexible agents over monolithic models.
Alibaba’s Qwen landscape shows a complex but clear trajectory: bigger is not always better. Agile, environment-aware agents are the real prize. Lin’s exit signals that the old model-focused thinking is giving way to something harder, messier—and far more useful.
Based on
- Qwen’s Former Lead on What Hybrid Thinking Got Wrong — and Why He Now Backs Agents — marktechpost.com
- Qwen’s Former Lead on What Hybrid Thinking Got Wrong — and Why He Now Backs Agents – The Future Tech — thefuturetech.co.uk
- Qwen’s former lead pivots from models to agents, citing hybrid thinking challenges · PulseAugur — pulseaugur.com
- Alibaba Unveils Qwen3.5: China’s Chatbot Race Shifts to AI Agents (2026) — rickcamerondesign.com
- The Frontier is Coming From Inside the House: Why Qwen 3.6 27B Just Made Your API Bill Obsolete – TechResider Submit AI Tool — techresider.com
