Whisper Large V3

Open speech recognition model commonly used for transcription and multilingual audio workflows.

Editorial review

Reviewed byOpenSourcesAI EditorialLast updatedAugust 2026SourcesExact model card

Model checkpoints, context windows, provider support, local runtime compatibility, and license terms can change quickly. Verify the exact model card before production or commercial use.

Best for

Builders adding local transcription, podcast processing, meeting notes, or audio translation.

Who should use it

Builders adding local transcription, podcast processing, meeting notes, or audio translation.
Builders who want local or self-hosted testing options.

Common workflows

Speech recognition, transcription, translation
audio workflows
transcription workflows

Deployment and hardware notes

~3.1 GB in fp16 fits any modern GPU comfortably; int8 through faster-whisper halves that and is the usual choice for batch jobs. CPU transcription works but runs slower than real time on most machines.

License and usage notes

Apache 2.0. Open source weights and code. Verify the exact model card and license terms for the checkpoint or hosted provider you use.

Strengths

Open source weights and code model option for Whisper workflows.
Builders adding local transcription, podcast processing, meeting notes, or audio translation.
Feed it 16 kHz mono audio. Transcription and speech-to-English translation are the same model with a different task token, so no second checkpoint is needed for translation.
Tracked as Audio in the OpenSourcesAI model directory.

Limitations

The slowest Whisper to decode, because the full 32-layer decoder runs for every output token. Long-audio jobs need chunking with overlap, and the model will confidently transcribe speech-like noise as words.
~3.1 GB in fp16 fits any modern GPU comfortably; int8 through faster-whisper halves that and is the usual choice for batch jobs. CPU transcription works but runs slower than real time on most machines.
Context window and limits: 30-second audio windows · 448-token cap per window.
Verify the exact model card, provider docs, license, and serving support before production use.

Local workflow notes

Feed it 16 kHz mono audio. Transcription and speech-to-English translation are the same model with a different task token, so no second checkpoint is needed for translation.

Local runtimes: Transformers, faster-whisper, whisper.cpp

Platforms: Windows, macOS, Linux

Transcription spec

Memory~3.1 GB in fp16 (1.5B parameters) · ~1.5 GB in int8Decoder depth32 layersLanguagesMultilingual (99 languages)Audio window30-second audio windows · 448-token cap per window

The accuracy reference for the family: a full 32-layer encoder and 32-layer decoder, and the only Whisper generation trained on 128 mel bins rather than 80. Every faster variant on this page is measured against it.

Sources to verify

Exact model card

Related resources

Continue with model source notes, local tools, and implementation guides related to this model.

Whisper family hub AI tools directory Model source notes Models directory Guides library

Hardware~3.1 GB in fp16 (1.5B parameters)RuntimeTransformers, faster-whisper, whisper.cppContext30-second audio windows · 448-token cap per windowLast updated2026

Exact model card →