whisper-medium-pl-v2-mlx-q8

whisper-medium-pl-v2-mlx-q8 is a 8-bit MLX affine (bits: 8, group_size: 32) Polish speech-to-text checkpoint derived from bardsai/whisper-medium-pl-v2 (24-layer encoder, 24-layer decoder, 80 mel bins, Common Voice 11 + FLEURS Polish fine-tune) for local transcription on Apple Silicon Metal. Upstream ships legacy PyTorch weights (3.06 GB FP32 PyTorch); this repository packages a validated, self-contained 861.8 MB (821.87 MiB) MLX safetensors bundle ready for mlx_whisper, LibraxisAI/mlx-batch-server, and vetcoders/x9bus.

Converted directly from bardsai/whisper-medium-pl-v2 into standard MLX 8-bit group-affine format (bits = 8, group_size = 32) across all 385 linear and vocabulary embedding projections while keeping encoder.conv1, encoder.conv2, decoder.positional_embedding, and all LayerNorm parameters in FP16. For tighter unified-memory budgets, see the companion 4-bit checkpoint LibraxisAI/whisper-medium-pl-v2-mlx-GPTQ-Hessian-q4 (482.9 MB).

Intended use

  • Local Polish speech-to-text dictation and batch audio transcription on Apple Silicon (M1 / M2 / M3 / M4)
  • Low-latency voice-to-agent and desktop dictation pipelines requiring compact unified-memory residency (861.8 MB weights)
  • Polish ASR evaluation and comparison against multilingual Whisper checkpoints

Out of scope

  • Safety-critical, medical, or legal decisions without domain expert review
  • Non-MLX runtimes (such as faster-whisper / CTranslate2 or whisper.cpp GGML), which use different tensor layouts
  • Unadapted English or multilingual transcription where a Polish-specialized fine-tune adapts foreign words phonetically into Polish orthography

Training and conversion metadata

Parameter Value
Repository LibraxisAI/whisper-medium-pl-v2-mlx-q8
Upstream checkpoint bardsai/whisper-medium-pl-v2 (derived from openai/whisper-medium)
Upstream Polish benchmark WER = 8.71%, CER = 2.41% on Mozilla Common Voice 11.0 (pl test split)
Task automatic-speech-recognition (language="pl")
Format MLX affine (bits: 8, group_size: 32)
Architecture whisper (d_model: 1024, n_audio_layer: 24, n_text_layer: 24, n_mels: 80, n_vocab: 51865)
weights.safetensors 861,792,281 bytes (861.8 MB / 821.87 MiB)
tokenizer.json 2,480,466 bytes
mel_filters.npz 4,271 bytes (mel_80 + mel_128)
config.json 342 bytes
Quantization scheme MLX group-affine Q8 (bits = 8, group_size = 32, 385 quantized matrices)
Unquantized tensors encoder.conv1, encoder.conv2, decoder.positional_embedding, LayerNorm weights/biases (F16), alignment_heads (I64)

Tested inference path

**Inference for this checkpoint has been tested with LibraxisAI/mlx-batch-server, vetcoders/x9bus (x9stt & codescribe-whisper-validate), and mlx-whisper.**
This is the recommended tested path for operator-controlled local inference on Apple Silicon.

Aspect Status
Tested runtimes LibraxisAI/mlx-batch-server, vetcoders/x9bus (x9stt), mlx_whisper
Target hardware Apple Silicon (M1 / M2 / M3 / M4)
Inference mode Local / self-hosted
Hugging Face Hosted Inference Disabled for this repository (inference: false)

Evaluation on Polish & mixed PL/EN speech (Apple Silicon Metal)

Measured via x9stt (CODESCRIBE_MODEL_PATH, native Metal backend, greedy decoding, language="pl", Silero VAD silence gating enabled, zero overlap with GPTQ calibration audio):

Checkpoint Weights Decoder layers Isolated numerals (5× "pięć", avg logprob) Continuous Polish speech excerpt Mixed PL/EN phrase ("…z computer use")
LibraxisAI/whisper-large-v3-turbo-mlx-q8 913.7 MB 4 Pięć pięć pięć pięć pięć (-0.22) Zastanawiamy, czy już się udało temu modelu wejść w rolę… (-0.31) …pracuje z computer use.
LibraxisAI/whisper-medium-pl-v2-mlx-q8 861.8 MB 24 pięć, pięć, pięć, pięć, pięć, pięć. (-0.13) Mamy zasadę, jakiejś, by już się udało temu modelu wejść w rolę i jest … (-0.19) …pierwsza aplikacja, przy której się pracuje z Computerius.
LibraxisAI/whisper-medium-pl-v2-mlx-GPTQ-Hessian-q4 482.9 MB 24 pięć, pięć, pięć, pięć, pięć, pięć. (-0.13) Zasada ja nie czy już się udało temu modelu wejść w rolę i jest tutaj z… (-0.20) …pierwsza aplikacja, przy której się pracuje z Computerius.

Usage

Python (mlx_whisper)

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="LibraxisAI/whisper-medium-pl-v2-mlx-q8",
    language="pl",
)
print(result["text"])

x9bus & x9stt CLI (macOS Apple Silicon)

hf download LibraxisAI/whisper-medium-pl-v2-mlx-q8 \
  --local-dir ~/.x9bus/models/whisper-medium-pl-v2-mlx-q8

CODESCRIBE_MODEL_PATH=~/.x9bus/models/whisper-medium-pl-v2-mlx-q8 \
  x9stt transcribe audio.wav -l pl --whisper

Limitations

  • Fine-tuned exclusively on Polish datasets (Common Voice 11.0 + Google FLEURS): delivers strong Polish orthography and inflection without clinical domain bias, but adapts foreign English technical terms toward Polish spelling; for mixed PL/EN developer dictation prefer LibraxisAI/whisper-large-v3-mlx-GPTQ-Hessian-q4 or LibraxisAI/whisper-large-v3-turbo-mlx-q8.
  • Purely silent or room-noise recordings without upstream Silero VAD gating can elicit standard Whisper sign-off hallucinations; keep VAD speech gating enabled for long recordings.
  • Memory footprint and decode latency depend on Apple Silicon generation, unified-memory bandwidth, and audio duration.

License

Released under Apache-2.0, matching upstream bardsai/whisper-medium-pl-v2.

Citation

@misc{libraxisai-whisper-medium-pl-v2-mlx-q8,
  title        = {whisper-medium-pl-v2-mlx-q8},
  author       = {LibraxisAI},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/LibraxisAI/whisper-medium-pl-v2-mlx-q8}},
  note         = {Polish MLX Whisper checkpoint published by LibraxisAI (derived from bardsai/whisper-medium-pl-v2)}
}

𝚅𝚒𝚋𝚎𝚌𝚛𝚊𝚏𝚝𝚎𝚍. with AI Agents by Vetcoders (c)2024-2026 LibraxisAI

Downloads last month
19
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for LibraxisAI/whisper-medium-pl-v2-mlx-q8

Finetuned
(2)
this model