The Qwen3 series, based on the Apple MLX framework Depth Optimization, has been announced by the general Quwen3 in Ali Baba, covering 32 official Qwen3 MLX models and a one-time full source. From Mac Pro, Mac Studio to Mac Mini, MacBook, to iPad and iPhone, all apples can easily deploy Qwen3 to achieve full-scale AI coverage.
The Qwen3 series model, which was launched in April 2025, is the most recent major AI model, using a hybrid structure (MoE), which supports 119 languages and dialects and is better performing than the former Qwen2.5. Thirty-two MLX models issued this time cover eight dimensions (from 0.6 B to 235 B parameters) and each model provides four quantitative precisions of 4bit, 6bit, 8bit and BF16 to meet the needs of different hardware resources.
The Apple MLX framework is known as the Optimized Open Source Machine Learning Tool for Apple Silicon Chip (M1/M2/M3/M4) for efficient training and deployment of large models. The Qwen3 Depth was optimized to achieve low delay and high performance reasoning on apple equipment. For example, the 4bit Quantification Model can operate on a smaller iPhone memory, while the BF16 high-precision model is suitable for high-performance devices such as Mac Studio.
In February 2024, Apple announced that it would work with Ali to develop the Chinese version of Apple Inteligence, Qwen3 as the first large-scale model of national production for a fully adapted MLX, paving the way for Apple AI to enter China. In response to a question from the Open Source Director, Lin Jun-hyun said: “This is a small update, but because of the number of models we have spent a lot of time testing with Mac Studio, hoping to help MLX users.”
The developers can access the model directly through the Hugging Face and the ModelScope platforms (HuggingFace Link: HuggingFace.co/Qwen3-MLX; ModelScope Link: ModelScope.cn/models) and run local AI applications on MacBook, iPad and even iPhone.
