Qwen3.5 27B

The Qwen3.5 27B is a dense vision-language model built to bring flagship-class capabilities into a more manageable footprint. Unlike the sparse Mixture-of-Experts variants in the Qwen3.5 series, this 27B model uses a dense architecture combining Gated Delta Networks with Feed Forward Networks, incorporating a linear attention mechanism to deliver fast response times while balancing inference speed and performance. Early fusion training on multimodal tokens gives it cross-generational parity with the Qwen3 base models, and it outperforms earlier Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks. The model supports 201 languages and dialects, extending global accessibility, while its native 262K context window can scale to approximately 1 million tokens for extended workflows.

The 27B model benefits from reinforcement learning scaled across million-agent environments with progressively complex task distributions, supporting robust real-world adaptability. This RL-driven approach cultivates capabilities across multimodal understanding and tool use, aligning the model's behavior with diverse production scenarios. Benchmarks place its overall capabilities on par with the much larger Qwen3.5-122B-A10B MoE model, making the 27B dense variant a compelling choice when compute efficiency matters without sacrificing core competencies in reasoning, coding, and multilingual generation. The open-weight availability and efficient inference profile position it well for developers seeking a capable, versatile foundation model for custom deployments or research applications.