whisper-large-v3-turbo-fp16-ov-npu

whisper-large-v3-turbo (OpenVINO IR, fp16) for Intel NPU

OpenVINO 2026 stateful export of openai/whisper-large-v3-turbo, suitable for openvino_genai.WhisperPipeline on Intel NPU (Panther Lake "AI Boost" / Lunar Lake / Meteor Lake).

Why this exists

Existing community NPU exports (e.g. FluidInference/whisper-large-v3-turbo-int4-ov-npu) were produced with an older OpenVINO that emits the legacy two-file decoder shape (openvino_decoder_model.xml + openvino_decoder_with_past_model.xml) without a beam_idx input. OpenVINO 2026's NPU plugin runs a StatefulToStateless transform that requires beam_idx, so loading those older IRs on NPU fails with:

Stateful models without `beam_idx` input are not supported in StatefulToStateless transformation

This repo holds a fresh export done with current optimum-intel (single-file stateful decoder + beam_idx input) so the NPU plugin in OV 2026 can load it directly.

How it was produced

optimum-cli export openvino \
    --model openai/whisper-large-v3-turbo \
    --task automatic-speech-recognition-with-past \
    --weight-format fp16 \
    whisper-large-v3-turbo-fp16-ov-npu

Versions used: openvino==2026.0.*, optimum-intel[openvino]>=1.20, transformers>=4.45.

Usage

import openvino_genai
pipe = openvino_genai.WhisperPipeline("./whisper-large-v3-turbo-fp16-ov-npu", "NPU")
text = str(pipe.generate(audio_float32_16k))

First-run NPU compile takes several minutes and populates ~/.cache/ze_intel_npu_cache. Subsequent runs are fast.

Notes

• fp16 weights, ~1.6 GB on disk.
• Inherits the Apache-2.0 license from the base model.