千问大模型：Qwen3.5-Flash来袭，三款中等规模模型全开源

Core Insights - Qwen 3.5 models have been officially open-sourced, including Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, and Qwen3.5-27B, showcasing significant performance improvements over previous larger models [1][2][12] Model Performance - Qwen3.5-35B-A3B outperforms the previous larger models Qwen3-235B-A22B-2507 and Qwen3-VL-235B-A22B, indicating that performance is now driven by architectural optimization, data quality enhancement, and reinforcement learning rather than just parameter scaling [1][3][13] - The Qwen3.5-122B-A10B and Qwen3.5-27B further narrow the performance gap between medium-scale models and cutting-edge models, particularly excelling in complex agent scenarios [1][3][13] Architectural Innovations - Qwen3.5 employs a hybrid attention mechanism combined with a highly sparse MoE architecture, trained on a larger scale of mixed text and visual tokens, achieving greater performance with fewer total and active parameters [3][10][15] - The new models have surpassed larger models in various authoritative benchmarks, including IFBench, GPQA, HMMT 25, MMMLU, BFCL v4, and SWE-bench Verified [3][10][15] Developer Accessibility - The Qwen3.5-27B model is designed for local deployment, featuring enhanced agent capabilities and native multimodal abilities, outperforming GPT-5 mini in multiple agent evaluations [4][16] - Qwen3.5-Flash API service is available on Alibaba Cloud, priced at 0.2 yuan per million tokens, offering high performance and cost-effectiveness for developers and enterprises [5][17] Community Support - All three models are available on platforms like Magic搭 and Hugging Face, along with the open-sourced Qwen3.5-35B-A3B-Base model to support community research, fine-tuning, and secondary development [7][19]