M87 Labs compresses Nvidia's Parakeet speech model to 178MB
Parakeet Redux runs at a claimed 113x realtime on eight CPU cores, with sharper multilingual scores and a measurable penalty in noisy audio.
By Ryan Merket · Published
Primary source: X
Why it matters
Parakeet Redux puts capable multilingual transcription inside a 178MB package, making local voice features more practical on CPUs and laptops while exposing the accuracy cost of extreme compression.
Vik Korrapati (@vikhyatk), co-founder of M87 Labs, released Parakeet Redux on September 21st, 2026, compressing Nvidia's 600-million-parameter speech recognition model from 1.2GB to 178MB by reducing its encoder weights to three possible values.
The result is a 1.58-bit speech-to-text model that M87 Labs says can transcribe audio at 113x realtime using eight x86 CPU cores. That means one second of processing handles 113 seconds of audio under the test configuration. On a MacBook Air with Apple's M2 chip, M87 Labs measured 38x realtime on the CPU and 43x using the integrated GPU.
Korrapati and his co-founder, identified by M87 Labs as Jay, started the San Francisco-based business in 2024 after working together at Amazon Web Services for eight years. Their better-known work is Moondream, a family of compact vision-language models. Parakeet Redux pushes the same small-model thesis into speech recognition and gives M87 Labs another workload for Photon, its local inference runtime.
A smaller Parakeet, with tradeoffs
Parakeet Redux is derived from Nvidia's Parakeet TDT 0.6B v3, a multilingual automatic speech recognition model released in August 2025. Nvidia's model supports 25 European languages, automatically detects the spoken language and can return word- or segment-level timestamps.
M87 Labs kept Parakeet's architecture and tokenizer while replacing each encoder weight with -1, 0 or 1. Its packed format stores five ternary values in each byte, according to the model repository. Korrapati said the smaller representation reduces the amount of weight data that processors must move while transcribing audio, addressing a memory-bandwidth bottleneck in CPU inference.
The current model card reports that Parakeet Redux reached 113x realtime on eight physical cores of an AMD EPYC 9575F processor. Under the same test, parakeet.cpp using 8-bit weights reached 45x, while two ONNX-based implementations reached 42x and 28x.
Those figures were produced by M87 Labs using Photon, its own inference engine. The comparison used LibriSpeech test-clean, processed one utterance at a time, with each competing runtime operating at its default settings. The 113x result is therefore a specific server-class benchmark, rather than a general speed guarantee for any eight-core desktop processor.
Compression holds up until the room gets noisy
M87 Labs' accuracy results show where the aggressive compression works and where it starts charging interest.
Across seven English datasets from the Open ASR Leaderboard, the original Parakeet recorded an average word error rate of 6.26%, compared with 6.55% for Parakeet Redux. Lower is better, leaving the compressed version 0.29 percentage points behind its base model.
Parakeet Redux performed better on the 25-language FLEURS benchmark, reducing average word error rate from 11.62% to 10.56%. The average conceals substantial differences between languages. Redux improved the reported results for Estonian, Latvian, Lithuanian, Maltese, Slovak and Slovene, while producing higher error rates for French, German, Polish, Russian and several others.
Long recordings also favored the compressed model in M87 Labs' tests. On 11 complete TED-LIUM talks lasting 10 to 20 minutes, Parakeet Redux recorded a 2.51% error rate, against 2.71% for Nvidia's original.
Noise is the clearest weakness. Across nine tests that mixed speech with background audio, Redux averaged 9.04% word error rate, compared with 6.72% for the base model. The model card says the compressed encoder is more likely to substitute similar-sounding words when the signal-to-noise ratio falls. Business-speech tests also favored the original, 6.15% to 6.96%.
The speed comparison carries an accuracy wrinkle of its own. Parakeet Redux posted a 1.94% error rate during the x86 performance test, while the slower parakeet.cpp implementation reached 1.51%. Developers choosing between the runtimes are trading memory and speed against transcription accuracy, rather than receiving a free 2.5x acceleration.
M87 Labs turns Photon into a broader local AI runtime
Parakeet Redux is available under the same CC BY 4.0 license as Nvidia's base model. Developers can run it locally through version 2.4.0 or later of the Moondream Python package, with CPU, Apple Metal and CUDA device options. The release includes support for sentence- and word-level timestamps, plus a voice-activity component that divides long recordings at pauses without a separate segmentation model.
For M87 Labs, the release broadens Photon beyond the computer-vision models on which Korrapati built Moondream. M87 Labs describes Photon as an inference engine for vision, audio and language models, tuned for low-latency workloads rather than large request queues. Speech recognition gives that pitch a second modality and a practical target: local transcription and voice applications running without a dedicated Nvidia GPU.
M87 Labs is backed by Felicis, Microsoft's M12 and GitHub fund, and Ascend. Parakeet Redux shows where Korrapati is directing part of that bet: smaller model files paired with a runtime designed around the hardware already sitting on a developer's desk. The benchmark tables also draw the boundary plainly. Ternary weights can cut a gigabyte-scale speech model down to 178MB, but difficult audio still exposes what the compression removed.