The MLX version of Qwen3.6-27B with support for multi-token prediction has been released

Developers introduced Qwen3.6-27B-MTPLX-Optimized — a specialized version of the language model, adapted for operation on devices with Apple Silicon architecture. The release is focused on using the native capabilities of the MLX framework to accelerate inference. The key feature of the model is support for native multi-token prediction (MTP) and speculative decoding. During tests, the system demonstrated up to 2.24x higher decoding speed when working with coding tasks. The use of MTP made it possible to achieve higher draft acceptance rates compared to standard implementations based on vLLM. The model is optimized for local launch on Mac, which allows for efficient use of the resources of Apple M-series graphics cores. Developers note that the MTPLX architecture provides a balance between generation quality and response delay, which is critical for interactive development tools. This release expands the capabilities of local inference of heavy models on consumer hardware. The transition to native MLX implementations makes it possible to reduce computational overhead and increase the overall token throughput per second without losing accuracy characteristic of quantized versions.