Accept the CrisperWhisper 2.0 model license

Converted weights and their Outputs are restricted to non-commercial use under the complete license in LICENSE.md. Commercial and production use require a separate license from nyra health GmbH.

By requesting access, you agree to the complete nyra health Non-Commercial Research License in LICENSE.md.

CrisperWhisper 2.0 Small for MLX

The small CrisperWhisper 2.0 checkpoint, converted to FP16 for Apple Silicon.

This conversion was built with MLX Porter, the open-source MLX porting toolkit from Amal David. It handled the checkpoint inspection, model route, source pins, tensor accounting, and parity plan. More details are at mlx-porter.pages.dev.

Here is the exact conversion you are downloading:

Source: nyralabs/CrisperWhisper2.0_small
Source revision: dbb990d5817659bfe2033835fb6e696ee18ebca0
Source weight SHA-256: 9da67d3597cb8344568535bff5c0ae72be697a728480275ef58a104bc2dfdf77
Converted weight SHA-256: bb660c455a9514a652b080f0fe2d667191528957703373db72370ca907557719
Dtype: FP16
Tensor accounting: 481 source tensors, 478 emitted, with every source tensor mapped or explained in WEIGHT_MAP.json
Snapshot accounting: source weights, config, tokenizer, processor, and license files are all hash-pinned in WEIGHT_MAP.json

Quick start

The inference code is in the separately MIT-licensed Amal-David/crisperwhisper-2-mlx repo:

pip install git+https://github.com/Amal-David/crisperwhisper-2-mlx.git@bd38e0020c3d22a6daa91ffc62f48be405030137
hf auth login
crisperwhisper-mlx audio.wav \
  --model amal-david/CrisperWhisper2.0_small-MLX \
  --mode verbatim \
  --word-timestamps

Version 0.2 supports:

verbatim and intended transcription
long audio through continuation chunks
word timestamps using CrisperWhisper's alignment path
Verbatimize conditioned decoding
guarded repetition repair and early-EOT recovery
CrisperWhisper-style hotword prompts
opt-in strict speculative verification

A few caveats

This standard small checkpoint was not trained for hotword boosting. The API accepts hotwords, but we are not making a hotword-accuracy claim.
Strict speculation is for correctness. Same-model drafting is not faster, and no faster compatible draft has been validated for this checkpoint.
Temperature fallback is not implemented.
Verbatimize handles one 30-second window and warns before truncating longer audio.
Early-EOT recovery is skipped with a warning when strict speculation is on.
The source-compatible full-window timestamp aligner can put a final word a few tens of milliseconds past the nominal clip duration.

Conversion details

The MLX graph uses deterministic sinusoidal encoder positions and a tied output projection. The two encoder_blank_head tensors are training-only, so the released inference and alignment paths do not use them. WEIGHT_MAP.json has the full mapping and hashes.

License

These converted weights are a Derivative Work of CrisperWhisper 2.0, so they keep the complete nyra health Non-Commercial Research License in LICENSE.md. The license also covers model Outputs. Commercial or production use needs a separate written commercial license from nyra health GmbH.

The inference code is MIT-licensed separately. That MIT license does not relicense these weights, the tokenizer/config files, or Outputs. This conversion is independent and is not affiliated with or endorsed by nyra health GmbH.

Downloads last month: 3

Safetensors

Model size

0.2B params

Tensor type

F16

MLX

Hardware compatibility

Quantized

Inference Providers NEW

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

Model tree for amal-david/CrisperWhisper2.0_small-MLX

Base model

nyralabs/CrisperWhisper2.0_small

Finetuned

(1)

this model