Small but Strong, Lightweight but Fast! Qwen3.5 Introduces Multiple Small-Sized Models Compatible with Consumer Graphics Cards

Tongyi Lab announced the official release of the latest small-sized version of its Qwen3.5 series, a new generation of large language models. The released models include four parameter scales: 0.8B, 2B, 4B, and 9B, aiming to reduce the application barriers of AI technology through extreme performance optimization, enabling low-cost and efficient deployment from edge devices to vertical scenarios.

According to the information, all models in this series are developed based on the unified Qwen3.5 base. Compared with large models that pursue extreme parameter sizes, these "small-size" members focus on "lightweight" and "high adaptability". The 0.8B and 2B models are specifically designed for edge devices, achieving extreme lightweight and millisecond-level fast response in environments such as smartphones and embedded hardware. The 4B model excels in multimodal capabilities and is considered an ideal choice for building lightweight agents (intelligent entities). Although the 9B model has a compact size, its actual performance is close to larger-scale models and can handle complex logical reasoning tasks.

To further embrace the developer ecosystem, Tongyi Lab announced that the series of models follows the Apache 2.0 license, making them open source and commercially usable. This means developers can freely perform LoRA or full fine-tuning, and only need common consumer-grade GPUs to start task adaptation. This move greatly reduces the time and cost for individual developers and small and medium enterprises to validate ideas and develop vertical applications.