Benchmarks

On-device performance for TheStage Apple SDK pipelines. Numbers below are from release builds on Apple M2 Max (NPU / ANE), macOS 26.2, measured 2026-07-28. Use them as guidance for relative speed and footprint — not as an SLA. Results move with chip, OS, thermals, and bundle revision.

Debug builds are much slower; always measure release.

Metric definitions

Metric

Meaning

TTFT

Prefill time to first token (LLM / ASR decoder)

TTFA

Streaming time-to-first-audio (TTS)

tok/s

Steady-state decode rate after warmup

rtf / rtfx

audio_seconds / wall_seconds (>1 = faster than realtime)

mem (MB)

Process footprint after the measured window

LLM (warmed, 256 new tokens, NPU)

Decode tok/s is nearly flat across prompt length; TTFT grows with prefill size (small → medium → long suite prompts).

Model

HF repo

tok/s (medium)

TTFT small → long (s)

mem (MB, medium)

Qwen3-0.6B

TheStageAI/Qwen3-0.6B

58.5

0.063 → 0.158

271

Gemma3-1B

TheStageAI/Gemma3-1B

83.1

0.058 → 0.112

163

LFM2.5-230M

TheStageAI/LFM2.5-230M

252

0.039 → 0.053

52

LFM2.5-350M

TheStageAI/LFM2.5-350M

178

0.038 → 0.060

61

TTS (warmed, NPU)

Bundle

HF repo

Batch rtf

Stream rtfx

TTFA (s)

mem (MB)

NeuTTS nano multilingual

TheStageAI/neutts-nano-multilingual

2.66

2.95

0.36

~215

Qwen3-TTS 12 Hz 0.6B

TheStageAI/Qwen3-TTS-12Hz-0.6B-Base

1.76

2.18

0.30

~423

Output is 24 kHz mono. Qwen3-TTS benches use official sampling defaults with a frame cap for stable runs.

ASR (warmed, NPU)

Primary shipping STT is Whisper. Qwen3-ASR is listed for comparison.

Bundle

HF repo

Audio

rtfx

decode tok/s

mem (MB)

thewhisper-large-v3-turbo

TheStageAI/thewhisper-large-v3-turbo

~2.6 s (say fox sentence)

16.7

233

~96

Qwen3-ASR 0.6B

TheStageAI/qwen3-asr-0.6b

8.88 s fixture

13.9

39.6

~342

Input is 16 kHz mono. Whisper shipping windows are 10 s.

Agent checklist

  • Quote device class + release build when citing numbers.

  • LLM: report tok/s and TTFT; they answer different questions.

  • TTS: prefer stream rtfx / TTFA for voice-agent UX; batch rtf for offline synth.

  • ASR: prefer rtfx for realtime; tok/s is decoder throughput only.