Sou Hu Cai Jing·2026-02-25 16:07Core Insights - Alibaba has launched the Qwen3.5 series of medium models, including Qwen3.5-122B-A10B, Qwen3.5-35B-A3B, and Qwen3.5-27B, which are open-source and designed for various applications in AI [2] - The new models demonstrate competitive performance in instruction following, reasoning, mathematics, multilingual knowledge, and agent tool usage, comparable to leading international models like GPT-5 mini and Claude Sonnet 4.5 [2] Model Performance - Qwen3.5-35B-A3B has outperformed larger models such as Qwen3-235B-A22B-2507, showcasing that smaller models can achieve superior performance on less expensive hardware [12][74] - The Qwen3.5-122B-A10B model excels in multi-modal, video, and multi-language scenarios, while Qwen3.5-35B-A3B focuses on deep reasoning and task scheduling capabilities [78] - Qwen3.5-27B is optimized for lightweight deployment, demonstrating high practical utility in programming and long-text processing [78] Cost Efficiency - Developers have reported significant cost savings, with potential monthly expenses dropping from €2000 for Gemini services to €50 using Qwen3.5-35B-A3B [12][74] - The Qwen3.5-Flash version offers a low cost of 0.2 yuan per million tokens, making it accessible for complex tasks and long documents [14] Market Position - The Qwen series has increased its market share in the enterprise-level model sector, with daily usage rising from 17.7% to 32.1%, solidifying its leading position [18][97] - The rapid release of the Qwen3.5 series models within a week of the Qwen3.5-397B-A17B launch indicates a strategic move to enhance product offerings and meet developer demands [97] Technical Advancements - Qwen3.5 models incorporate a unified visual language foundation and an efficient mixed architecture, enhancing throughput and reducing latency [96] - The models support 201 languages and dialects, ensuring global deployment and cultural understanding [96]