Historical catalog and model evidence. Runtime support, ratings, UI behavior, and commands below describe the implementation at the recorded date. Retired benchmark commands are retained as provenance, not instructions. See the desktop guide for current supported artifacts and features; these records do not claim Electron performance.
This report records Textify’s promotion evidence for the exact
mlx-community/whisper-large-v3-turbo model requested for the native MLX path.
It does not generalize the English measurements to other languages.
| File | Bytes | SHA-256 |
|---|---|---|
config.json |
268 | b34fc29e4e11e0a25e812775dd67f4dd16fc2c8eb43d28ae25ff7d660ecb6379 |
weights.safetensors |
1,613,977,612 | 951ed3fc1203e6a62467abb2144a96ce7eafca8fa77e3704fdb8635ff3e7f8a6 |
tokenizer.json |
2,710,337 | 297b13372ac43916285644fb9687add3cc62ee2a1adb60da3dc25cc94c1871fd |
tokenizer_config.json |
282,843 | 844b642c73a91359722f47b35705f7174686df33d252695d8572cf9ac03a6389 |
special_tokens_map.json |
2,186 | baea4ea09372eb4fca86b4e4346139fd73cb807d5087e9de0948e971739c3e74 |
added_tokens.json |
34,648 | 3c51f66c4c21f9e126970078f11ae77a78c74aee8df606ee9daba86e467108e0 |
vocab.json |
1,036,558 | e2aa043ef015641d363d8288e7c241c85e36a5c761fb303598e0710233344387 |
merges.txt |
493,869 | 2df2990a395e35e8dfbc7511e08c12d56018d8d04691e0133e5d63b21e154dc6 |
normalizer.json |
52,666 | bf1c507dc8724ca9cf9903640dacfb69dae2f00edee4f21ceba106a7392f26dd |
generation_config.json |
3,772 | cce11bfe3aaa6ae9e072ea2637caaec8795e68d9b67e655a5af16ee509681a4c |
The 1,618,594,759-byte directory uses the MLX checkpoint from revision
a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb. The eight tokenizer/generation
files come from the original openai/whisper-large-v3-turbo revision
41f01f3fe87f28c78e2fbf8b568835947dd65ed9. Installing those files in the
managed directory prevents the pinned Swift runtime from using its mutable
network tokenizer fallback. The original checkpoint is MIT licensed.
Textify used MLX Audio Swift revision
d302a5c6080d2bb97bae38c7418f82abb76013b6 and MLX Swift 0.31.3 revision
61b9e011e09a62b489f6bd647958f1555bdf2896, both under MIT. The required
3,135,962-byte mlx.metallib targets macOS 14.0 and has SHA-256
cffe8fbfa9cfb794f1d920ff187016f823a555f7814cede99026699a936b92c7.
dev-clean-2 utterances in Corpus/openslr31.json,
totaling 210 reference words under CC BY 4.0.| Word errors / words | Aggregate WER | Release-to-final median | p95 | Median RTF | Peak resident memory |
|---|---|---|---|---|---|
| 7 / 210 | 3.33% | 321.5 ms | 416 ms | 0.0458 | 1,748,041,728 bytes |
Per-clip release-to-final values, in corpus order, were 270, 287, 273, 258, 373, 378, 380, 416, 296, and 347 ms. Warm process model loads measured 192-207 ms with a 196 ms median; warmups measured 232-256 ms with a 232 ms median. The first shader-preparation smoke recorded a 210 ms load and 1.873 s warmup before caches were populated.
Textify’s adapter requires the exact ten-file managed directory, an actual MLX Metal backend, explicit English, 16 kHz mono audio, and a 60-second maximum. It rejects automatic language detection, non-English selection, CPU fallback, and incomplete tokenizer directories. Digital and near-digital silence returns no speech without invoking the model.
The downloaded files matched the signed catalog sizes and SHA-256 values. The
native runtime loaded the exact directory with Textify’s packaged
mlx.metallib and produced A man said to the universe, Sir, I exist. at 0
WER on the first fixed clip. The opt-in signed-catalog integration test covers
atomic installation and the same local runtime route.
The exact artifact is promoted as a separate Experimental choice. It meets Textify’s local-only, immutable-artifact, Metal-routing, English accuracy, and interactive latency gates. Compared with the existing 5-bit whisper.cpp Turbo entry, it improved the narrow English result from 4.29% to 3.33% WER and the median from 367 ms to 321.5 ms, but increases the download from 574 MB to 1.62 GB and peak process memory from about 700 MB to 1.75 GB. Both choices remain in the picker so that tradeoff stays explicit.