3x faster speeds on MLX | Qwen 3.8 27B | Native MTP Speculative Decoding On Apple Silicon With No External Drafter.
metal mtp mlx inference-engine apple-silicon local-ai qwen speculative-decoding speculative-sampling openai-compatible qwen3-next anthropic-compatible native-mtp mtplx
-
Updated
Aug 20, 2026 - Python