oMLX 0.6.1 tunes Qwen3.8 inference

oMLX is becoming a serious local inference stack for Apple Silicon, optimizing the hardware instead of treating Macs as secondary GPU platforms.

• –Dual-ANE/GPU prefill targets long-context workloads, where prompt processing is often the biggest bottleneck.
• –Lightning MTP improvements make speculative decoding more useful for interactive coding-agent sessions.
• –SSD-backed KV caching remains the standout differentiator for repeated, shifting prompts.
• –The release reinforces oMLX’s focus on Qwen3.8 and large-memory Macs rather than broad cross-platform support.
• –Experimental hardware paths may require careful validation across chip generations and model quantizations.