Editorial illustration for Liquid AI Releases LiquidCodec, a 74M-Parameter Streaming Audio Codec
Liquid AI Releases LiquidCodec Audio Codec
Liquid AI put out LiquidCodec on Hugging Face this week, a 74-million-parameter audio codec built to run underneath its LFM3-Audio-2B and LFM3-Omni-2B models. The job of a codec like this is narrow but important: turn speech into tokens a language model can chew on, then turn those tokens back into audio that sounds right. LiquidCodec does both, and according to Liquid AI's own numbers it does the encoding part fast enough to run on a single CPU thread, 361 times faster than real time on an Apple M4 Max laptop.
It's a direct swap for Kyutai's Mimi, the codec Liquid AI used in its earlier LFM2-Audio-1.5B model. Mimi carries 96.2 million parameters; LiquidCodec trims that down while still beating it, X-Codec 2 and NanoCodec on ViSQOL and mel distance scores across eight test sets, per the model card. The architecture borrows blocks from Liquid AI's own LFM3 design for both the encoder and decoder, with a fully causal encoder built for streaming. The quantization step is where things get more unusual.
Liquid AI has released LiquidCodec, the audio codec behind its LFM3-Audio-2B and LFM3-Omni-2B models. The LiquidCodec audio codec is a ~74M-parameter streaming speech tokenizer. It turns speech into tokens a language model can predict, then converts those tokens back into audio.
Why this matters
For anyone building voice agents or real-time speech systems, the RTF numbers are the story. Mimi runs at 0.0361 encode RTF, NanoCodec at 0.2145, X-Codec 2 at 0.7359. LiquidCodec's 0.0028 isn't a marginal gain over those, it's a different category of speed, and that gap compounds fast once you're chaining a codec with an LLM in a live pipeline.
At 74M parameters, this is small enough to run on-device or alongside a larger model without eating your compute budget. Liquid AI designed it specifically to feed LFM3-Audio-2B and LFM3-Omni-2B, so the real test is whether token quality holds up once you're past benchmark tables and into messy production audio: accents, noise, overlapping speech. Weights are on Hugging Face, so developers can check that themselves rather than take the speed claims on faith.
We'd watch for independent latency tests on actual hardware, not just the RTF ratios Liquid published, and for whether other audio-LLM builders start treating 12.5 Hz frame rates as the new baseline rather than an outlier.
Related Reading
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Axiom Math raises USD 200M Series A led by Menlo to keep AI slop out of math
From Archives Nvidia CEO Exit, Hugging Face Deal in Doubt
From Archives OpenAI quarterly revenue reportedly tops USD 10B as losses shrink
Common Questions Answered
What is LiquidCodec and what are its primary functions?
LiquidCodec is a 74-million-parameter streaming audio codec developed by Liquid AI that converts speech into tokens that language models can process, then converts those tokens back into audio. It serves as the underlying codec for Liquid AI's LFM3-Audio-2B and LFM3-Omni-2B models, enabling efficient audio-to-text-to-audio conversion in real-time applications.
How does LiquidCodec's encoding speed compare to other audio codecs?
LiquidCodec achieves an RTF (Real-Time Factor) of 0.0028, which represents a significant performance advantage over competing codecs like Mimi (0.0361), NanoCodec (0.2145), and X-Codec 2 (0.7359). According to Liquid AI's measurements, LiquidCodec can encode audio 361 times faster than real time on a single CPU thread, placing it in a different category of speed compared to existing solutions.
Why is LiquidCodec's efficiency important for voice agents and real-time speech systems?
The extreme speed of LiquidCodec's encoding (0.0028 RTF) becomes critical when chaining it with language models in live pipelines for voice agents. At only 74M parameters, LiquidCodec is small enough to run on-device or alongside larger models without consuming significant compute resources, making real-time speech processing practical and efficient.
What models does LiquidCodec support?
LiquidCodec is designed to work as the audio codec backend for Liquid AI's LFM3-Audio-2B and LFM3-Omni-2B models. These models leverage LiquidCodec's tokenization capabilities to process and generate audio in real-time applications.
Where can developers access LiquidCodec?
Liquid AI released LiquidCodec on Hugging Face, making it publicly available for developers building voice agents and real-time speech systems. This open release allows the community to integrate the codec into their own audio processing pipelines.
Further Reading
- LiquidCodec: High-Fidelity Audio Tokenization Optimized for Speed - Liquid AI
- Build specialized multimodal models with LFM3 - Hugging Face
- LFM3-Audio-2B Model Card - Hugging Face
- StreamSE: Neural Audio Codec-based Streaming Speech Enhancement - arXiv
