vasanth009/parakeet-redux-mlx

Parakeet Redux in a custom MLX format

This is an experimental weight-only export for a standalone custom MLX Redux implementation. It repacks the trained Moondream Parakeet Redux weights. It is not a new training run or a quantization of NVIDIA's full-precision model.

This is not a drop-in Transformers AutoModel or MLX Audio checkpoint. No inference code, Python modules, custom AutoModel registration, Photon files, Kestrel files, or compiled runtime artifacts are distributed here. The source config and ternary manifest describe the original export; conversion.json defines this checkpoint's different storage format. A separately reviewed custom loader, feature extractor, FastConformer, predictor, joint network, and TDT decoder are required. Runtime-code licensing remains unresolved. The weight license does not grant rights to any separately distributed runtime.

Attribution and license

NVIDIA created parakeet-tdt-0.6b-v3, the underlying architecture and original model. Moondream created the trained ternary Redux derivative, including its voice-activity head. The upstream card credits NVIDIA and states that Redux uses CC BY 4.0, the same license as the original.

The authoritative Redux model card at revision 2bf128600aac4b16946f7ed8372e56117fe5e23b supplies that attribution and license notice. Its raw file SHA256 is 930055e6bac7d75114e719850ffaffc8d327e73b4295518c38c81399acd63e46. The NVIDIA model card independently identifies CC BY 4.0. This conversion consumes the pinned Redux export, not a separately downloaded NVIDIA checkpoint.

The weights and accompanying model metadata retain Creative Commons Attribution 4.0 International. LICENSE contains the complete unmodified official English legal text. Retain the source attribution and indicate these packing changes when redistributing. The material is provided as-is, without warranties, as detailed in Section 5 of the license. Neither NVIDIA nor Moondream endorses this conversion.

The upstream card also links the Redux and Ultra release post, Parakeet Ultra, and Photon. These links describe upstream work, not compatibility or performance claims for this artifact. Upstream Photon benchmark numbers do not establish this implementation's performance.

Changes and storage contract

The change from the pinned Redux export is lossless integer packing. There is no calibration, retraining, or requantization. All trained ternary values and original FP16 per-row, per-128-element scales are preserved. Nonternary tensors retain their original values and dtypes. Config, tokenizer, and source ternary metadata are copied byte for byte.

The original format stores five little-endian base-3 codes per byte. This export stores sixteen little-endian two-bit codes per uint32. Only codes 0, 1, and 2 are used. With group size 128, weight = scale * code - scale reconstructs the trained values. The custom implementation supplies affine biases as negative scales at use time; they are not separate saved tensors. Original .qweight entries become .packed entries. The 264 packed matrices and their scales, together with the other 465 tensors, produce 993 saved tensors. Tensor digests use logical little-endian C-order bytes.

The model remains a 24-layer FastConformer with a two-layer LSTM predictor and TDT head. The implementation's input contract is mono 16 kHz audio with 128 mel bins. Its tokenizer has 8,193 entries including blank ID 8192, pad ID 2, and duration choices 0, 1, 2, 3, and 4. This description is not an included inference API.

Artifact	SHA256
Pinned upstream weights	78ec25733ee0d0c1586d1346fc86db9d0c2e436e3a8ab1d32a82d1bb8f848d21
Converted model.safetensors, 207858007 bytes	03f41729b69807deec2d3bb7dc7d45f5c0285d781fd598c4c4a848079cf1a0d9
Independent packed_tensors.json	cd74909c621e6c5b1349bd604a8a2d51cfbcfc2d07fa71189f5240a6e11c6f73

source-model.json records the exact upstream revision, source file sizes and hashes, attribution links, and legal-text digest. conversion.json pins converted files. release-manifest.json pins every other released file. packed_tensors.json contains independently established source-derived tensor hashes, shapes, and dtypes. Keep a trusted copy of the release pins outside the downloaded folder. A manifest that an attacker can replace is not a signature or independent evidence of authenticity.

Recorded parity and known failures

These are reported results from the pinned local evaluation, not a new evaluation of this staging operation. parity-summary.json is an allowlisted projection of that report. It contains aggregates, evidence hashes, and failure descriptions, with no audio, raw transcripts, or user filesystem paths.

Check	Recorded result
Independent fixed-window transcripts	41/41 valid cases exactly matched the reference
Expected too-short rejection	Matched for 1 additional case
Reference-segment token, duration, and frame paths	44/44 exactly matched in each category
Aligned vocabulary and duration argmax	2046/2046 matched in each domain
Strict tensor acceptance	Failed, with 4 checks failing on 2 quiet cases
Original implementation tests	75 passed and 1 quiet-feature test failed, out of 76

quiet-pcm16 failed normalized-feature allclose, subsampling normalized L2, and layer-0 normalized L2. quiet-low failed normalized-feature allclose. The separate quiet floating-point unit probe failed with normalized-feature maximum absolute difference 0.3542853. These failures remain unsuppressed. Near-floor preprocessing rounding can amplify small differences through normalization. Lossless packing does not establish identical floating-point execution. No identical-FP, bit-identical inference, complete numerical parity, or production-readiness claim is made.

The corpus has 42 cases and about 286 seconds. It contains English read speech, synthetic European-language speech, boundary and silence probes, quiet probes, and concatenated long speech. It is smoke coverage, not evidence of accuracy across all 25 upstream languages, noisy conversation, code-switching, natural meetings, or hour-scale recordings. Transcript agreement with another engine is not ground-truth accuracy. Both engines had known-reference fixed-window WER 7/728 and CER 10/3294 on this small corpus.

The recorded candidate used MLX 0.31.2 on Apple M5 with 32 GiB and macOS 26.6.2 arm64. The reference used Moondream 2.6.1, Kestrel 0.9.1, and Torch 2.14.1. Its canonical target was MPS with FP16 activations. Native VAD-driven long-form segmentation, batching, native Swift integration, and compatibility with other MLX versions remain unvalidated. The candidate used fixed 28-second windows with 2-second overlap rather than the reference's native VAD segmentation.

Package boundary

This package contains only model.safetensors, config.json, tokenizer.json, ternary.json, conversion.json, packed_tensors.json, source-model.json, parity-summary.json, README.md, LICENSE, and release-manifest.json. Preparation and verification scripts stay outside the model package. No private workspace report, transcript, corpus audio, dependency environment, or runtime implementation is included. Publication and downstream application integration require separate approval.