This is a custom Q8 quantization of OpenAI's Whisper large-v3-turbo, produced for the Madi macOS transcription app (a pure Zig + Metal engine for Apple Silicon). The Madi app downloads this file on first run instead of bundling it, keeping the app ~37 MB.

The weights are pre-quantized exactly as the Madi engine quantizes them in memory at load time — per-32-element blocks, scale = max(|w|) / 127 stored as f16, q = clamp(round_half_away(w / scale), -127, 127) as int8. Linear/projection weights are stored as K.qs (int8) + K.scales (f16); layernorm, bias, conv, and positional-embedding tensors are copied in F16 unchanged.

Because the offline quantization is identical to the engine's in-memory path, loading this Q8 file produces transcripts that are byte-identical to loading the original F16 weights — the 1.86× size reduction is free of quality loss. This is a verified contract (F16-load vs Q8-load output equality), not an approximation.

This is not a general-purpose checkpoint. The safetensors file uses a layout specific to the Madi Metal engine (.qs / .scales keys). It is not compatible with 🤗 Transformers, whisper.cpp (GGML/GGUF), faster-whisper, or other Whisper runtimes. Use one of those projects' own quantized checkpoints for those runtimes.

Released under the MIT License, inherited from the base model openai/whisper-large-v3-turbo (also MIT). When redistributing, please retain attribution to OpenAI's Whisper.