antflydb/whisper-large-v3-turbo

whisper-large-v3-turbo (GGUF, Q8_0)

openai/whisper-large-v3-turbo exported to a single GGUF bundle for the Antfly inference runtime. The tokenizer, BPE merges, and model configuration (including the 128 mel bins) are embedded, so the file runs on its own.

Quantization policy: the decoder projection layers, the encoder, and the token table are Q8_0; the position tables stay F32. On the same clips this bundle produces transcripts identical to the published f16 checkpoint, at 829 MB instead of 1.5 GB, with a faster encoder and prompt block on Apple silicon.

Use

antfly inference pull antflydb/whisper-large-v3-turbo:gguf:Q8_0 --tasks transcribe

Then pass "model": "antflydb/whisper-large-v3-turbo" to /ai/v1/transcribe, /ai/v1/dictate, or a transcription session.

Reproduce

antfly inference pull openai/whisper-large-v3-turbo --tasks transcribe
antfly inference export ~/.antfly/inference/models/openai/whisper-large-v3-turbo --target gguf --format q8_0 \
  --output whisper-large-v3-turbo.Q8_0.gguf

License

The weights are derived from OpenAI's Whisper, released under the MIT license; this export keeps that license.