[ncylich]

Summary

Fills the README benchmark table for the two devices measured so far — Mac M4 Pro and Galaxy S26. README-only change; numbers come from the existing cactus test engine suites (no new tooling).

Device	Gemma4 Text (pf/dec tps)	Gemma4 Vision (lat/dec)	Parakeet (lat)	RAM
Mac M4 Pro	324 / 39	1.2s / 48 (NPU)	1.2s	1385 MB
Galaxy S26	248 / 21	17s / 16	4.2s	4024 MB

How the numbers were produced

All via existing suites on CQ4 bundles (cactus download apple bundle on Mac for the NPU vision encoder):

cactus test --component engine --suite llm --model google/gemma-4-E2B-it   # Gemma4 Text (test_1k_context: ~1k prefill / 100 decode)
cactus test --component engine --suite vlm --model google/gemma-4-E2B-it   # Gemma4 Vision (256px)
cactus test --component engine --suite stt --transcription-model <parakeet bundle>  # Parakeet (20s test.wav)

Notes for reviewers

• NPU vs CPU. Per the table preamble, VLM/Transcribe are meant to use NPU prefill. NPU (Apple ANE) is Mac-only.
  • Mac Gemma4 Vision is NPU — the cactus download apple bundle ships vision_encoder.mlpackage; prefill jumps to ~240 tps / ~1.2s latency vs ~84 tps / ~3.3s on CPU.
  • Mac Parakeet is CPU. There is no published Parakeet-TDT apple bundle, and a local --npu transpile fails — coremltools can't convert the TDT audio encoder (Unsupported fx node _assert_tensor_metadata). So this cell is CPU until that conversion is supported.
  • Galaxy S26 is CPU for all three (no ANE path on Android).
• Parakeet is a TDT transducer, so decode_tps isn't meaningful — the Parakeet column reports latency for 20s of audio (RTF ≈ 16× on Mac, ≈ 4.8× on Samsung).
• LLM number uses the existing test_1k_context suite (≈1221 prefill tokens, 100 decode) rather than a new exact-1000 fixture, per the request to reuse existing tests.
• Device label "Galaxy S26" follows the existing table; the connected unit reports model SM-S942U1.
• The "Parakeet 1.1B" column header still reads 1.1B while the model is 0.6B (pre-existing; left untouched).

[ncylich]

Consolidated into #729, which records these Mac M4 Pro and Galaxy S26 numbers alongside the Parakeet NPU work — Parakeet transcription on Mac now runs on the Neural Engine (1.2s CPU → 0.2s NPU).