• Sources: primary, discussion
  • Summary: The post, published 2026-10-02, presents one 16.9 MB file against 145.3 MB for Whisper base and 41.9 MB for Moonshine tiny v2, CPU only with no dependencies, seven languages, and 30 seconds of 16 kHz mono audio per pass with word timestamps and encoder embeddings. On ten seconds of audio on an Apple M4 Pro CPU it reports 11.1 ms to first token and 1,319 decoded tokens per second against 73.2 ms and 266 for Whisper base, and word error rate ahead on LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22 and the FLEURS average but behind Whisper base on TED-LIUM, AMI and the MLS average. These are the vendor's own figures and not independently reproduced, and the stated method is scoring with the Whisper normalizers over 86,174 utterances, each model on its official runtime at its defaults, decontamination verified by comparing audio checksums and speaker IDs across every reported test set, and missing bars marked as benchmarks the other authors never published.
  • Why it matters: A stated benchmark method, including decontamination checked by audio checksum and speaker ID across every reported test set, is what separates a size and latency claim a reader can test from one that has to be taken on trust.
  • Follow-up: Independent reproduction of the word error rate and latency figures against Whisper base and Moonshine tiny v2.