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 Parakeet RNNT 1.1B model requested from the supplied Artificial Analysis comparison. It does not generalize the English measurements to other languages.
| File | Bytes | SHA-256 |
|---|---|---|
config.json |
37,318 | 1d84911a9139a9f56a43cfc8d46400e125bcade8a39508d604492fe291c8f40b |
model.safetensors |
4,282,246,596 | d2c370b3728c4c3b318d814574e87da144a9484ce6b3c4457a6ed5889bb007e1 |
tokenizer.model |
259,162 | 8f279c3951a1c280a99a108d07b429bd98c03c94ecd0270748bd7affca9c0817 |
tokenizer.vocab |
11,383 | dc8f48909c2d3a0374f45b7478226d26a7de16bbc5334448a8e989f4538384d1 |
vocab.txt |
5,301 | 27277f2a7176399c7acc178e394cec5e6a575657db9aae9e9e609cb004e9f350 |
The 4,282,559,760-byte directory came from
mlx-community/parakeet-rnnt-1.1b revision
7f399a0d3442123deae9194e71f5c984b2879efa. The model is CC-BY-4.0.
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 |
|---|---|---|---|---|---|
| 4 / 210 | 1.90% | 128.5 ms | 241 ms | 0.0182 | 4,456,595,456 bytes |
Per-clip release-to-final values, in corpus order, were 85, 86, 82, 73, 192, 143, 175, 208, 114, and 241 ms. Warm process model loads measured 235-243 ms; the benchmark’s first shader-packaging smoke recorded a 899 ms load and 1.724 s warmup before caches were populated, so those distributions are not treated as a clean cold-start comparison.
MLX Swift’s command-line SwiftPM path does not compile Metal shaders. A build
without the MLX library failed before inference; substituting Whisper’s
default.metallib loaded a library but failed on the missing
layer_normfloat32 kernel. Textify now ships a distinct, hash-verified
mlx.metallib colocated with the executable and release verification checks
the kernel name to prevent regression.
A fresh install-shaped directory was populated with the five exact files,
every SHA-256 was rechecked, and the native runtime produced
a man said to the universe sir i exist at 0 WER while reporting the required
MLX Metal backend.
The exact artifact is promoted as Experimental. It meets Textify’s local-only, immutable-artifact, Metal-routing, English quality, and latency gates. Its 4.28 GB download and 4.46 GB peak process footprint are substantially larger than the existing choices, so the picker requires at least 16 GB unified memory and advertises only English until more fixed-corpus evidence exists.