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 Canary-Qwen 2.5B model requested from the supplied Artificial Analysis comparison. It does not generalize the English measurements to other languages.
| File | Bytes | SHA-256 |
|---|---|---|
canary-qwen-2.5b-Q4_K_M.gguf |
1,737,575,808 | db5162229d6fa22597d06a613bd9b543eddb3ee02e6afc5e759120fde02bebf7 |
The file came from handy-computer/canary-qwen-2.5b-gguf revision
3370d4e2f28cc70eea79dfc9f2f43fb91eef3163. The original NVIDIA source is
pinned at revision b1469e1bba1cfe140205529c79c434ca47180960. The model and
conversion are CC-BY-4.0, its Qwen component is Apache-2.0, and Textify’s
transcribe.cpp 0.1.3 runtime is pinned to
5a5a49664a8ea1f0e5b3be1dfc544730d1b62561 under MIT.
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 |
|---|---|---|---|---|---|
| 2 / 210 | 0.95% | 255.5 ms | 373 ms | 0.0312 | 3,163,389,952 bytes |
Per-clip release-to-final values, in corpus order, were 143, 216, 143, 123, 323, 295, 327, 373, 171, and 308 ms. Warm process model loads measured 490-504 ms and warmups 80-99 ms. The first native process also spent 6.555 seconds loading the embedded ggml Metal library; later process loads used the compiled Metal cache, so that first initialization is disclosed separately rather than included in release-to-final latency.
Textify’s adapter rejects CPU fallback, automatic language detection,
non-English language selection, and audio beyond the model’s 40-second quality
window. It also bypasses the autoregressive runtime for digital or near-digital
silence. The pinned dynamic library is loaded RTLD_LOCAL so its ggml symbols
cannot collide with Textify’s separate whisper.cpp copy.
An opt-in native regression loaded the exact Q4_K_M file through the shipped
runtime directory, verified the Metal route, and produced
A man said to the universe, Sir, I exist under 700 ms. The signed-catalog
integration covers atomic download/install and the same runtime route when its
network-heavy opt-in gate is enabled.
The exact artifact is promoted as Accurate. It meets Textify’s local-only, immutable-artifact, Metal-routing, English accuracy, and interactive latency gates. Its 1.74 GB download, 3.16 GB peak process footprint, English-only coverage, and 40-second cap are disclosed directly in the picker.