The quantized model Qwen3.8-Flash-Next has been released for use on Mac

Developers presented Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-3.3bpw — a specialized version of the language model, optimized for running on Apple devices with Apple Silicon architecture. The release is targeted at users with systems having 64 GB of RAM. The key feature of the build was quantization to 3.3 bits per weight (bpw). With these parameters, the model demonstrates 85.6% top-1 accuracy in accuracy tests. Using this version requires 52 GB of resident memory, which allows for effective use of Mac hardware resources for local calculations. The architecture supports working with a context window of up to 64 thousand tokens. During testing, the model showed a result of 8 points on the internal performance evaluation scale, achieving this indicator in 2 hours of data processing. The toolkit is based on the MLX framework, which provides native support for acceleration on Apple M-series chips. This allows running heavy LLM without the need to use cloud GPUs, while maintaining high generation accuracy. This release expands the capabilities of local launch of large models for developers and researchers working in the macOS ecosystem. Optimizing weights to 3.3 bits allows the model to be placed in the memory of devices with 64 GB RAM, providing a balance between resource consumption and response quality.