The mixture-of-experts model has 26.6 billion total and approximately 4.6 billion active parameters. It supports transcription style instructions, context hints, multilingual ASR and recognition of challenging entities such as addresses and code symbols.
Why it matters
More accurate entity transcription could improve voice-driven agents and developer tools.
A community MLX 4-bit conversion of SOV-Transcribe-1 enables experimentation on sufficiently equipped Apple Silicon Macs; its documented validation is narrow.
Salesforce's DFlash draft model claims around 2–3× acceleration for some single-request speech workloads, subject to model, hardware and recording length.