Junyang Lin is the founder of Pragmatik Labs and a former technical leader of Alibaba’s Qwen open-weight models, which helped make advanced language, coding, reasoning, and multimodal capabilities accessible to developers worldwide. His Shanghai-based company is building agents for digital and physical worlds: systems that can understand their environments, use tools, and complete extended tasks.
Lin studied at Peking University and joined Alibaba in 2019. His early work included KOBE, a personalized product-description system deployed on Taobao, combining Transformer-based generation with product attributes, audience information, and external knowledge. He later co-authored the original Qwen technical report and the Qwen3 technical report, helping develop a model family spanning dense and mixture-of-experts architectures, mobile-scale deployments, multilingual applications, and demanding reasoning workloads.
- Controllable reasoning budgets. Qwen3 combines immediate responses and deliberate multistep reasoning within one model. Developers can cap reasoning tokens to balance accuracy, latency, and computational cost, making sophisticated inference adjustable for specific products.
- Open models built for practical deployment. Lin championed downloadable weights, multiple model sizes, quantized formats, permissive licensing, and compatibility with established inference frameworks. Qwen3 supports 119 languages and dialects, extending these capabilities across local devices, private infrastructure, and international applications.
- Agents that learn from environmental feedback. Lin’s approach combines reasoning, tool use, observed outcomes, and continued planning. His applications include desktop organization, website generation, and research workflows; WebWorld, which he co-authored, provides simulated browser environments for training web agents.
- Multimodal systems designed for action. His work spans vision-language models, image and video understanding, audio interaction, and multilingual streaming speech synthesis. He has emphasized that adding modalities must preserve language, reasoning, and agent performance, while visual understanding enables computer-use tasks unavailable to text-only systems.
Lin left Qwen in March 2026 and announced Pragmatik Labs that August, identifying Gaorong Ventures, HSG, Tencent, and the Shanghai Future Industry Fund as supporters. His new venture advances the direction he outlined at AI Engineer World’s Fair 2025: training systems that improve through reinforcement learning, richer perception, longer context, and sustained interaction with their surroundings.
