File	Quantization	Description
`ggml-model-f16.bin`	float16	Full precision (no quantization) — highest quality
`ggml-model-q8_0.bin`	int8	8-bit quantization — minimal quality loss
`ggml-model-q5_0.bin`	int5	5-bit quantization — good quality/size tradeoff
`ggml-model-q4_0.bin`	int4	4-bit quantization — smallest size, fastest

Training

Fine-tuned openai/whisper-large-v3 (1.55B params) on YODAS-Granary Italian (asr_only ~87k + ast 200k cap, conversational YouTube audio).

See LocalAI-io/whisper-large-v3-it-yodas-only for the full safetensors model and detailed WER results.

Usage

whisper.cpp

bash

# Download a quant
huggingface-cli download LocalAI-io/whisper-large-v3-it-yodas-only-ggml ggml-model-q5_0.bin --local-dir .

# Run
./whisper-cli -m ggml-model-q5_0.bin -f audio.wav -l it

whisper.cpp Python bindings (pywhispercpp)

python

from pywhispercpp.model import Model

model = Model("ggml-model-q5_0.bin", language="it")
segments = model.transcribe("audio.wav")
for seg in segments:
    print(seg.text)

LocalAI

yaml

# In your LocalAI model config
name: whisper-large-v3-it-yodas-only
backend: whisper
parameters:
  model: ggml-model-q5_0.bin

LocalAI-io/whisper-large-v3-it-yodas-only-ggml

whisper-large-v3-it-yodas-only-ggml

Files

Training

Usage

whisper.cpp

whisper.cpp Python bindings (pywhispercpp)

LocalAI

Links

LocalAI-io/
whisper-large-v3-it-yodas-only-ggml