I recommand using the 1.5B model. of course, you can choose another model depending on you device, but it must meet these requirements:
- It must be a DeepSeek-R1 distill model.
- It must be in MLX format.
- Your device must be capable of running it (with sufficient memory).
You can find the models on huggingface
Examples:
- mlx-community/deepseek-r1-distill-qwen-1.5b
- mlx-community/DeepSeek-R1-Distill-Qwen-7B-4bit
- mlx-community/DeepSeek-R1-Distill-Llama-8B
我建议使用 1.5B 模型。当然,你也可以根据你的设备选择其他模型,但必须满足以下要求:
- 必须是 DeepSeek-R1 distill 模型。
- 必须是 MLX 格式。
- 你的设备必须能够运行(内存够大)。
你可以在 Hugging Face 上找到这些模型。
例如: