Orukeet is a 25-language speech recognizer from Oruk AI, with co-authors from Stanford, Cambridge, OpenWhispr and Hoid, adapted from NVIDIA Parakeet TDT 0.6B v3; the GitHub repository was created September 9, 2026, alongside weights on Hugging Face (oruk/orukeet) and an arXiv preprint (2609.10054). The model keeps Parakeet's 627,008,134 parameters, 24-layer FastConformer encoder, token-and-duration transducer and tokenizer, but each encoder block's 1,024 nine-tap temporal depthwise filters are re-fit: all 24,576 filters are fitted to a Gabor function, and the 12,288 with the lowest normalized squared error are frozen as fixed kernels, leaving 626,897,542 scalar parameters trainable. A final r3 adaptation pass runs over the same 2,939 LibriSpeech test-other recordings later used for checkpoint selection. The README reports Orukeet beating Parakeet on 61 of 74 tested splits: LibriSpeech test-clean 1.46% vs. 1.53% WER, test-other 2.86% vs. 3.14% (test-other is the adaptation/selection split, not a held-out result), FLEURS English 3.82% vs. 4.28%, and pooled FLEURS across 25 languages 9.85% vs. 11.01%, a 10.6% relative reduction. Released artifacts, a 2.51 GB NeMo checkpoint, a 714 MB native Q8 GGUF (the default format), a 1.30 GB native F16 GGUF, and a 487 MB ONNX INT8 archive, derive from the same pinned r3 checkpoint; GGUF files run on a pinned NVIDIA NeMo-Speech.cpp SDK, not llama.cpp. The install --device auto command downloads a prebuilt Metal, CUDA, Vulkan, or CPU runtime; since v0.1.1, Apple silicon gets an Orukeet-hosted Metal SDK built from patches in the repo. Hoid's Metal kernel and attention/cache patches cut warm Q8 median latency 42.4% (154.2 to 88.9 ms) on an M4 Pro in a 24-clip patch benchmark, which has not been rerun on the packaged SDK. OpenWhispr 1.10.0 ships Orukeet as its recommended local model over sherpa-onnx, and TapTalk 0.4.0 offers an Orukeet Core ML engine, tagged as a preview release, though TapTalk's live typing still uses base Parakeet and only completed recordings run through Orukeet. Weights are cached locally and transcription then runs offline. Code is MIT licensed; the r3 weights and fitted Gabor kernels are CC BY-SA 4.0, retaining NVIDIA's original CC BY 4.0 attribution for the Parakeet foundation model.