Ming-TTS inference ported to Zig/MLX hits 18x speedup on M4/M5 chips
s_busso · reddit · 2026-10-31
Frustrated with Ming-TTS 3.0's default speed, developer sbusso ported inference to Zig/MLX, achieving up to 18x faster TTS on M4/M5 chips, open-sourced as ming-tts-serve.
Benchmarks (latency, lower is better):
- Official PyTorch 2.6.0 + MPS: 1.841 (EN) / 1.907 (ZH) / 1.951 (voice clone) — baseline 1x
- mlx-audio 0.6.3: 8x
- ming-tts-serve plain MLX graph (MINGFUSED=0): 14x
- ming-tts-serve fused kernels (default): 18x (0.104 EN / 0.106 ZH / 0.112 clone)
Repo: github.com/sbusso/ming-tts-serve
More from Infra
- Claude Code background agents burn ~$4.80/hour just writing status lines, dev calculates — arpit_bhayani · 2026-10-31
- Two weeks of agent logs: INT4 quantization triples Qwen 27B's reasoning loops at long context — Old_Ad_6033 · 2026-10-31
- YOLOv2 Face Detection Hits 30fps on a K210 Chip via Custom C++ Firmware — alexcovo_eth · 2026-10-31
- Playable Environments: Millisecond Cloud Branching as the Next Test Bed for Agents — matei_zaharia · 2026-10-31
- Apple silicon is so overpowered only local LLM runners and 8K editors notice — vivekhaldar · 2026-10-31
- Wafer hits #3 on OpenRouter by daily token volume at 1.9T — ycombinator · 2026-10-31