Core AI model zoo

item	iPhone run 1 (fresh install, cold)	iPhone run 2 (relaunch, warm)	Mac A1 Release (cold for this app)	Mac A2 Release (warm)
run / status / pass	20260927-030040 / done / True	20260927-030429 / done / True	mac-20260927-025503 / done / True	mac-20260927-025652 / done / True
device	iPhone19,2 V63AP, iOS 27.0.0 (24A437), arch h19p	iPhone19,2 V63AP, iOS 27.0.0 (24A437), arch h19p	arm64 Mac16,9, macOS 27.0.0 (26A428), arch h16c	arm64 Mac16,9, macOS 27.0.0 (26A428), arch h16c
build	Release, testable True	Release, testable True	Release, testable True	Release, testable True
GPU lock (Mac)	-	-	taken by this run (flock; file 0 bytes, last written 2026-09-27 02:52:26; GPU utilization [56, 0, 0, 0, 0])	taken by this run (flock; file 0 bytes, last written 2026-09-27 02:56:52; GPU utilization [98, 0, 0, 0, 0])
assets: files / missing / md5 differ	325 / 0 / 0	325 / 0 / 0	325 / 0 / 0	325 / 0 / 0
load1 encoder alone (s)	2.46	0.10	2.30	0.11
load1 model = decoder + cached encoder (s)	2.82	0.90	1.44	0.55
load1 encoder warm (s)	0.14	0.15	0.11	0.12
load1 decoder ≈ model − encoder warm (s)	2.67	0.75	1.33	0.43
load1 encoder + decoder ≈ (s)	5.14	0.85	3.62	0.54
Core AI cache before → after load1 (MB)	0 → 1564	1564 → 1564	0 → 1564	1564 → 1564
warmup first transcription (ms, clip)	2256 (zh, exact)	552 (zh, exact)	221 (zh, exact)	216 (zh, exact)
load2 model, same process (s)	0.44	0.44	0.55	0.56
e2e clips done / errors	155 / 0	155 / 0	155 / 0	155 / 0
oracle exact	150	150	150	150
knife-edge (< 0.1)	5	5	5	5
exact or knife-edge	155	155	155	155
mismatch ≥ 0.1	0	0	0	0
ids == Python engine	155	155	155	155
ids == Mac Swift (r3)	155	155	155	155
text == oracle / Python / Mac Swift	150 / 155 / 155	150 / 155 / 155	150 / 155 / 155	150 / 155 / 155
cap (512)	0	0	0	0
front end max	Δ	vs NumPy (clip)	4.93e-04 (ja_jp_1670)	4.93e-04 (ja_jp_1670)	4.93e-04 (ja_jp_1670)	4.93e-04 (ja_jp_1670)
e2e audio / transcription wall (s)	1836 / 132.1	1836 / 141.7	1836 / 39.5	1836 / 39.2
e2e RTF median / p90	0.0745 / 0.0937	0.0807 / 0.0994	0.0221 / 0.0280	0.0220 / 0.0280
e2e RTF worst (clip)	0.1370 (ko)	0.1350 (ko)	0.0434 (ko)	0.0435 (ko)
e2e RTF median, clips started in the first 20 s / after	0.0780 (28) / 0.0745	0.0781 (28) / 0.0811	0.0234 (82) / 0.0215	0.0234 (83) / 0.0214
e2e median front end / encoder / prefill (ms)	2.2 / 146.1 / 392.8	2.4 / 225.1 / 417.2	1.4 / 40.2 / 123.7	1.3 / 40.2 / 122.6
e2e median decode (ms/token)	8.17	9.41	2.79	2.77
bench clip, runs, wait for nominal	ja_jp_1719, 5 after 1 warm-up, waited 46 s (fair → nominal)	ja_jp_1719, 5 after 1 warm-up, waited 302 s (serious → serious)	ja_jp_1719, 5 after 1 warm-up, waited 0 s (nominal → nominal)	ja_jp_1719, 5 after 1 warm-up, waited 0 s (nominal → nominal)
bench RTF median / p90	0.0665 / 0.0667	0.0667 / 0.0676	0.0218 / 0.0218	0.0216 / 0.0216
bench wall median (ms)	906.3	908.8	296.5	294.0
bench encoder / prefill (ms), decode (ms/token)	150.5 / 415.6, 8.18	151.0 / 416.0, 8.16	42.0 / 136.9, 2.80	42.0 / 135.6, 2.77
bench timed runs end at (s)	5.5	5.5	1.8	1.8
bench thermal start → end	nominal → nominal	serious → fair	nominal → nominal	nominal → nominal
peak footprint over the run (MB)	385	411	727	737
least available memory (MB)	3157	3133	-	-
e2e thermal start / ~20 s / end	nominal (0 s) / nominal (20 s) / fair (134 s)	nominal (0 s) / nominal (20 s) / serious (143 s)	nominal (0 s) / nominal (20 s) / nominal (40 s)	nominal (0 s) / nominal (20 s) / nominal (40 s)
thermal at launch → end of run	nominal → nominal	nominal → fair	nominal → nominal	nominal → nominal
battery at launch	80 % charging	80 % charging	-100 % unknown	-100 % unknown
elapsed (s)	197	456	50	47

iPhone run 1 (fresh install, cold): /Users/majimadaisuke/code/coreai/coreai-models-community-funasr-wt/apps/FunASRGate/_work/device_runs/20260927-030040/result.json
iPhone run 2 (relaunch, warm): /Users/majimadaisuke/code/coreai/coreai-models-community-funasr-wt/apps/FunASRGate/_work/device_runs/20260927-030429/result.json
Mac A1 Release (cold for this app): /Users/majimadaisuke/code/coreai/coreai-models-community-funasr-wt/apps/FunASRGate/_work/mac_runs/mac-20260927-025503/result.json
Mac A2 Release (warm): /Users/majimadaisuke/code/coreai/coreai-models-community-funasr-wt/apps/FunASRGate/_work/mac_runs/mac-20260927-025652/result.json

Testability control (Mac, paired per clip: the public-API build without -enable-testing / the testable build; run-to-run lines for scale)

A1 testable -> B1 public:
paired clips 155: wall B/A median 0.9933, mean 0.9934, min 0.9759, max 1.0075; sum B / sum A 0.9932; text equal 155/155
bench wall median A 296.483584 ms, B 293.42775000000006 ms
A2 testable -> B1 public:
paired clips 155: wall B/A median 0.9995, mean 0.9993, min 0.9813, max 1.0156; sum B / sum A 0.9992; text equal 155/155
bench wall median A 294.004166 ms, B 293.42775000000006 ms
A2 testable -> B2 public:
paired clips 155: wall B/A median 0.9992, mean 0.9993, min 0.9861, max 1.0112; sum B / sum A 0.9994; text equal 155/155
bench wall median A 294.004166 ms, B 294.58641600000004 ms
A1 testable -> A2 testable (run to run):
paired clips 155: wall B/A median 0.9939, mean 0.9941, min 0.9843, max 1.0100; sum B / sum A 0.9940; text equal 155/155
bench wall median A 296.483584 ms, B 294.004166 ms
B1 public -> B2 public (run to run):
paired clips 155: wall B/A median 1.0002, mean 1.0001, min 0.9872, max 1.0137; sum B / sum A 1.0002; text equal 155/155
bench wall median A 293.42775000000006 ms, B 294.58641600000004 ms
mac-20260927-025503: app priority min 31 (31 = foreground user process; App Nap = 4), demoted samples 0 , load avg start/end 17.7/28.2
mac-20260927-025557-public: app priority min 31 (31 = foreground user process; App Nap = 4), demoted samples 0 , load avg start/end 28.2/15.8
mac-20260927-025652: app priority min 31 (31 = foreground user process; App Nap = 4), demoted samples 0 , load avg start/end 15.8/11.3
mac-20260927-025747-public: app priority min 31 (31 = foreground user process; App Nap = 4), demoted samples 0 , load avg start/end 11.3/7.9

Phone vs Mac, and the phone’s speed over the e2e stage

iPhone run 1 vs Mac A2 per clip: waveform md5 equal 155/155, Swift front-end features md5 equal 155/155, generated ids equal 155/155; iPhone run 1 vs run 2 ids equal 155/155
iPhone run 1 thermal every 20 s: [(0, 'nominal'), (20, 'nominal'), (40, 'nominal'), (60, 'nominal'), (80, 'nominal'), (100, 'fair'), (120, 'fair'), (134, 'fair')]
  e2e   0- 20 s: 28 clips, wall iPhone/Mac x3.13, encoder 146 ms, decode 8.16 ms/token (medians)
  e2e  20- 40 s: 26 clips, wall iPhone/Mac x3.12, encoder 146 ms, decode 8.17 ms/token (medians)
  e2e  40- 60 s: 26 clips, wall iPhone/Mac x3.12, encoder 146 ms, decode 8.16 ms/token (medians)
  e2e  60- 80 s: 27 clips, wall iPhone/Mac x3.10, encoder 146 ms, decode 8.14 ms/token (medians)
  e2e  80-100 s: 19 clips, wall iPhone/Mac x3.62, encoder 227 ms, decode 8.81 ms/token (medians)
  e2e 100-120 s: 19 clips, wall iPhone/Mac x3.93, encoder 231 ms, decode 10.04 ms/token (medians)
  e2e 120-140 s: 10 clips, wall iPhone/Mac x4.08, encoder 232 ms, decode 10.75 ms/token (medians)
iPhone run 2 thermal every 20 s: [(0, 'nominal'), (20, 'nominal'), (40, 'fair'), (60, 'serious'), (80, 'serious'), (100, 'serious'), (120, 'serious'), (140, 'serious'), (143, 'serious')]
  e2e   0- 20 s: 28 clips, wall iPhone/Mac x3.13, encoder 147 ms, decode 8.17 ms/token (medians)
  e2e  20- 40 s: 26 clips, wall iPhone/Mac x3.14, encoder 147 ms, decode 8.16 ms/token (medians)
  e2e  40- 60 s: 24 clips, wall iPhone/Mac x3.16, encoder 148 ms, decode 8.23 ms/token (medians)
  e2e  60- 80 s: 22 clips, wall iPhone/Mac x3.94, encoder 233 ms, decode 9.87 ms/token (medians)
  e2e  80-100 s: 18 clips, wall iPhone/Mac x3.92, encoder 234 ms, decode 9.94 ms/token (medians)
  e2e 100-120 s: 18 clips, wall iPhone/Mac x3.94, encoder 234 ms, decode 10.11 ms/token (medians)
  e2e 120-140 s: 17 clips, wall iPhone/Mac x4.06, encoder 234 ms, decode 10.49 ms/token (medians)
  e2e 140-160 s:  2 clips, wall iPhone/Mac x4.23, encoder 236 ms, decode 11.17 ms/token (medians)

This site is open source. Improve this page.