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 Cohere Transcribe model requested from the supplied Artificial Analysis comparison. It does not generalize the English measurements to other languages.
| File | Bytes | SHA-256 |
|---|---|---|
config.json |
4,336 | 206113921b2b75ad43461e9d2d4cd361405be9acaf29aa3ec1ed751c16347bb6 |
model.safetensors |
2,418,031,831 | bd1edbe982f47d22e64ba5723a4204b43ca93ea703b18946160d44ec26c83ab9 |
tokenizer.model |
492,827 | 6d21e6a83b2d0d3e1241a7817e4bef8eb63bcb7cfe4a2675af9a35ff3bbf0e14 |
tokenizer_config.json |
48,141 | 0dfeb3eeba07bccaa1b4bf78f3135ad3059acf8d18f681675832b285ac0035b0 |
The 2,418,577,135-byte directory came from
beshkenadze/cohere-transcribe-03-2026-mlx-8bit revision
d1f843476f84846e6fe7aa58a6033f17882f0ec9. The official model and conversion
are Apache-2.0. 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 was compiled from the nine generated
entry points in that pinned MLX Swift checkout, 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 |
|---|---|---|---|---|---|
| 6 / 210 | 2.86% | 121.5 ms | 304 ms | 0.0179 | 2,522,726,400 bytes |
Per-clip release-to-final values, in corpus order, were 81, 89, 83, 82, 137, 140, 281, 210, 106, and 304 ms. Warm process model loads measured 192-204 ms with a 196.5 ms median; warmups measured 71-93 ms with a 71.5 ms median. The first shader-preparation smoke recorded a 228 ms load and 1.090 s warmup before caches were populated, so that warmup is disclosed separately.
Textify’s adapter requires the exact four-file model directory, an actual MLX Metal backend, explicit English, and 16 kHz mono audio. It rejects automatic language detection, non-English selection, CPU fallback, and audio beyond 30 seconds. Digital and near-digital silence returns no speech without invoking the model.
All four downloaded files matched their immutable catalog sizes and SHA-256
values. The native runtime loaded those exact files with Textify’s packaged,
signed mlx.metallib and produced
A man said to the universe, Sir, I exist. at 0 WER on the first fixed clip.
The signed-catalog integration covers atomic installation and the same local
runtime route behind its network-heavy opt-in gate.
The exact artifact is promoted as Experimental. It meets Textify’s local-only, immutable-artifact, Metal-routing, English accuracy, and interactive latency gates. Its 2.42 GB download, 2.53 GB peak process footprint, English-only coverage, new conversion, and 30-second recording cap are disclosed directly in the picker.