GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Omar10lfc
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Omar10lfc
Model Tree
openai/whisper-large-v3-turbo
Input Modalities
Output Modalities
Supported Functionality
All models were evaluated on the canonical Mozilla Common Voice 17.0 Arabic test set using the standardized normalization protocol:
| Architecture | Parameters | Common Voice 17.0 Test WER | PyTorch FP16 Throughput | CTranslate2 INT8 Throughput | Hugging Face (FP16) | Faster-Whisper (CT2 INT8) |
|---|---|---|---|---|---|---|
| Whisper-Large-v3 | 1,550M | 12.51% | 5.0x realtime | 18.0x realtime | Omar10lfc/whisper-large-v3-arabic | Omar10lfc/whisper-large-v3-arabic-ct2-int8 |
| Whisper-Large-v3-Turbo | 809M | 14.15% | 8.5x realtime | 17.5x realtime | Omar10lfc/whisper-large-v3-turbo-arabic | Omar10lfc/whisper-large-v3-turbo-arabic-ct2-int8 |
| Whisper-Medium (Stage 2) | 769M | 18.16% | 4.5x realtime | 8.8x realtime | Omar10lfc/whisper-medium-arabic-stage2 | Omar10lfc/whisper-medium-arabic-stage2-ct2-int8 |
| Whisper-Small | 244M | 20.61% | 12.4x realtime | 28.5x realtime | Omar10lfc/whisper-small-arabic | Omar10lfc/whisper-small-arabic-ct2-int8 |
openai/whisper-large-v3-turbo (809M parameters, 32 encoder layers, 4 decoder layers)ar)Does 8-bit quantization degrade accuracy? Our empirical benchmark across the full Mozilla Common Voice Arabic evaluation suite demonstrates quantization parity—the Word Error Rate (WER) difference between unquantized FP16 and CTranslate2 Static INT8 is statistically negligible (≤0.5 to 0.8 percentage points), while reducing disk size and VRAM by ~50% and accelerating inference by up to 3.6×:
| Model Architecture | Precision Tier | Model File Size | Measured WER | Throughput (NVIDIA T4) | Quality Delta |
|---|---|---|---|---|---|
| Whisper-Medium | Vanilla PyTorch FP16 | 1,531 MB | 17.83% | 6.7× realtime | Baseline |
| CTranslate2 INT8 | 770 MB (−49.7%) | 18.35% | 8.8× realtime | +0.52 pp (statistically negligible) | |
| Whisper-Large-v3 | Vanilla PyTorch FP16 | 3,091 MB | 16.80% | 5.0× realtime | Baseline |
| CTranslate2 INT8 | 1,556 MB (−49.7%) | 17.63% | 18.0× realtime | +0.83 pp (3.6× speedup) | |
| Whisper-Large-v3-Turbo | Vanilla PyTorch FP16 | 1,621 MB | 18.46% | 8.5× realtime | Baseline |
| CTranslate2 INT8 | 819 MB (−49.5%) | 18.26% | 17.5× realtime | −0.20 pp (INT8 parity) |
*-ct2-int8): Recommended for production APIs, serverless containers (Modal, RunPod, Hugging Face Spaces), and edge environments to cut cold-start load times and memory by half while maximizing throughput.transformers pipeline integration.python
import torchfrom transformers import AutoModelForConditionalGeneration, AutoProcessor, pipeline model_id = "Omar10lfc/whisper-large-v3-turbo-arabic"processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForConditionalGeneration.from_pretrained( model_id, torch_dtype=torch.float16, device_map="cuda") pipe = pipeline( "automatic-speech-recognition", model=model, tokenizer=processor.tokenizer, feature_extractor=processor.feature_extractor, generate_kwargs={"language": "arabic", "task": "transcribe"}) result = pipe("path/to/arabic_audio.mp3")print("Transcription:", result["text"])import torchfrom transformers import AutoModelForConditionalGeneration, AutoProcessor, pipeline model_id = "Omar10lfc/whisper-large-v3-turbo-arabic"processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForConditionalGeneration.from_pretrained( model_id, torch_dtype=torch.float16, device_map="cuda") pipe = pipeline( "automatic-speech-recognition", model=model, tokenizer=processor.tokenizer, feature_extractor=processor.feature_extractor, generate_kwargs={"language": "arabic", "task": "transcribe"}) result = pipe("path/to/arabic_audio.mp3")print("Transcription:", result["text"])