--- type: "post" id: "liquid-ai-released-liquidcodec-a-74m-param-streaming-40a5e4" slug: "liquid-ai-released-liquidcodec-a-74m-param-streaming-40a5e4" title: "Liquid AI released LiquidCodec, a ~74M-param streaming audio codec for LFM3-Audio-2B and LFM3-Omni-2B." description: "Liquid AI has published LiquidCodec, a compact, streaming codec that turns speech into discrete tokens and reconstructs it as audio." retrieval_nugget: "Liquid AI has published LiquidCodec, a compact, streaming codec that turns speech into discrete tokens and reconstructs it as audio. The roughly 74-million-parameter component powers LFM3-Audio-2B and LFM3-Omni-2B; its low reported CPU real-time factors and quantized decoder make it relevant to teams building speech systems, although the published comparisons are Liquid AI’s own evaluations." published_at: "2026-11-13" updated_at: "2026-11-13" record_date: "2026-11-13" date_kind: "published_at" topics: ["models","multimodal","inference"] entities: ["Liquid AI","LiquidCodec","LFM3-Audio-2B","LFM3-Omni-2B","Transformers","llama.cpp","Qwen3-1.7B"] related_patterns: ["generative-media-becomes-application-infrastructure"] novelty_label: "new_fact" what_changed: "Liquid AI released LiquidCodec as a standalone streaming codec, with a roughly 74M-parameter model and reported fast encoding, high-fidelity reconstruction, and a compact decoder option." prior_refs: ["https://newruntime.com/posts/lfm3-audio-2b-supports-18-languages-tool-calls-8c0d35/"] source_url: "https://huggingface.co/LiquidAI/Liquid-Codec" source_urls: ["https://huggingface.co/LiquidAI/Liquid-Codec"] schema_version: "newruntime-agent-readable-v0.2" stable_id: "post:liquid-ai-released-liquidcodec-a-74m-param-streaming-40a5e4" status: "published" visuals: [] editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]} routes: {"html":"https://newruntime.com/posts/liquid-ai-released-liquidcodec-a-74m-param-streaming-40a5e4/","markdown":"https://newruntime.com/posts/liquid-ai-released-liquidcodec-a-74m-param-streaming-40a5e4.md","json":"https://newruntime.com/posts/liquid-ai-released-liquidcodec-a-74m-param-streaming-40a5e4.json"} markdown_url: "https://newruntime.com/posts/liquid-ai-released-liquidcodec-a-74m-param-streaming-40a5e4.md" --- # Liquid AI released LiquidCodec, a ~74M-param streaming audio codec for LFM3-Audio-2B and LFM3-Omni-2B. ## Retrieval answer Liquid AI has published LiquidCodec, a compact, streaming codec that turns speech into discrete tokens and reconstructs it as audio. The roughly 74-million-parameter component powers LFM3-Audio-2B and LFM3-Omni-2B; its low reported CPU real-time factors and quantized decoder make it relevant to teams building speech systems, although the published comparisons are Liquid AI’s own evaluations. ## Summary Liquid AI has published LiquidCodec, a compact, streaming codec that turns speech into discrete tokens and reconstructs it as audio. The roughly 74-million-parameter component powers LFM3-Audio-2B and LFM3-Omni-2B; its low reported CPU real-time factors and quantized decoder make it relevant to teams building speech systems, although the published comparisons are Liquid AI’s own evaluations. ## What happened Liquid AI’s Hugging Face model card describes LiquidCodec as an audio codec that compresses speech into discrete tokens and converts those tokens back into audio. It is the audio tokenizer behind LFM3-Audio-2B and LFM3-Omni-2B. The card lists approximately 74 million parameters, a symmetric, fully causal LFM3 encoder and decoder, and grouped finite scalar quantization (GFSQ). Its specified input and output are 24 kHz mono speech. It produces 12 tokens per frame at a frame rate of 12.5 Hz, with a bitrate of approximately 1.65 kbps. Both encoding and decoding support streaming. The design choice Liquid AI emphasizes is that GFSQ codebooks have no sequential dependencies: each group is quantized independently, allowing a speech model to predict all codebooks for a frame in parallel. The encoder mirrors the decoder and is fully causal. The card says the decoder predicts spectral coefficients directly, avoiding progressive upsampling with learned transposed convolutions. Together, these choices are intended to support real-time, low-latency processing. The repository includes BF16 weights for Transformers, with examples for loading the model through `AutoModel` and encoding and decoding waveforms. It also shows how to decode successive frames through a streaming cache. Liquid AI reports speed measurements for single-threaded FP32 inference on an Apple M4 Max. In that test, LiquidCodec’s encoding real-time factor was 0.0028 and its decoding factor was 0.0115; the card translates these to about 361 times real time for encoding and 87 times for decoding. It compares LiquidCodec with X-Codec 2, NanoCodec and Mimi, reporting lower factors for LiquidCodec in the displayed table. These results describe that specific setup, not every device or deployment. The card says a decoder-only quantization-aware training approach lets the decoder run in Q4_0 with a 3.5-times smaller footprint and minimal quality loss. Separate quantized codec weights are not published in this repository; the card says quantized decoders ship inside LFM3-Audio-2B-GGUF and LFM3-Omni-2B-GGUF for on-device inference with llama.cpp. The model card also reports reconstruction and text-to-speech evaluations across eight test sets. Liquid AI says LiquidCodec achieved the best ViSQOL and lowest mel distance on all eight reconstruction sets, including with a Q4_0 decoder. In matched TTS experiments, models trained on LiquidCodec tokens had the highest UTMOSv2 on all eight sets, and the best WER/CER on four. The TTS comparison used identical data, architecture, compute and training order across tokenizers; the card reports a second experiment with a Qwen3-1.7B backbone as a consistency check. LiquidCodec was trained and evaluated on speech using 264,000 hours across 107 languages. The listed license is LFM Open License v1.0. The card explicitly says quality on music, singing and environmental sound has not been evaluated. ## What the archive adds New Runtime covered LFM3-Audio-2B’s language support, tool calls, roughly 50 ms time to first token and reported LiquidCodec speed on [October 31, 2026](https://newruntime.com/posts/lfm3-audio-2b-supports-18-languages-tool-calls-8c0d35/). The codec repository now gives the component’s architecture, model-card access instructions, detailed comparisons and evaluation setup. This is a new technical fact on a known LFM3 audio story, not a restatement of the model’s user-facing capabilities. ## Trend signal LiquidCodec supports the pattern that generative media becomes application infrastructure when systems can compose it as a controllable component, rather than treating generation as a one-off artifact. The codec makes audio tokens, streaming and reconstruction part of the model stack; see [Generative media becomes application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/). This release is evidence of an infrastructure direction, not proof that the same component is suitable for every audio workload. ## Why now The repository positions the codec as the audio tokenizer for LFM3-Audio-2B and LFM3-Omni-2B, and describes quantized decoder weights bundled with their GGUF versions for on-device inference. That links the codec’s publication to the model family’s practical deployment paths, especially streaming and local use. The model card does not state a publication date or explain a broader launch timetable. ## Connections to other events New Runtime’s earlier coverage described LFM3 as an open-weight family spanning text, vision, audio, omni and decision models ([October 27, 2026](https://newruntime.com/posts/liquid-ai-launches-lfm3-an-open-weight-2b-d00d7d/)). Its later coverage noted that llama.cpp v0.8.0 added LFM3 support and experimental streaming audio, including realtime and TTS endpoints ([November 1, 2026](https://newruntime.com/posts/llama-cpp-v0-8-0-adds-liquidai-lfm3-a547b8/)). LiquidCodec adds detail about the audio representation and the codec weights bundled for local inference; it does not establish that every repository example works through those server endpoints. ## Why it matters A speech application’s model is only one part of its latency, quality and deployment profile. Its tokenizer and decoder determine how audio is represented, how quickly frames can be encoded or reconstructed, and how much memory a local runtime may need. LiquidCodec gives builders a concrete component to evaluate alongside an LFM3 audio model, with stated streaming behavior, token rate, bitrate and comparisons across reconstruction and TTS tasks. The reported CPU speed and smaller Q4_0 decoder footprint could make the codec worth testing in latency-sensitive or edge-oriented prototypes. But the decision should be workload-specific: the headline speed figures come from one CPU configuration, while the quality tables concern defined speech test sets and producer-run experiments. Teams should compare their own languages, microphones, noise conditions and latency targets, and separately establish whether the license and bundled model weights fit their deployment. In particular, the card’s speech-only evaluation does not support assuming equivalent results for music, singing or environmental audio. ## Limits and uncertainty The speed results are reported by Liquid AI for FP32, single-threaded inference on an Apple M4 Max; the model card does not establish performance across other CPUs, GPUs, mobile devices or production workloads. The quality tables are company-reported evaluations, not independent replications. The card specifies that music, singing and environmental sound quality has not been evaluated. It also does not quantify application-level latency, memory use across target devices, or how results vary across real-world recording conditions. The repository lists an LFM Open License v1.0, but teams still need to review the terms for their intended use. No separate quantized codec files are published in this repository. ## Skill to build now Run a small codec bake-off on a representative speech sample before integrating it. Load the BF16 repository model through Transformers, measure encode and decode time against audio duration on the hardware you plan to ship, and compare reconstruction quality with your current codec. If your deployment target is llama.cpp, test the bundled Q4_0 decoder in the relevant GGUF model rather than assuming the Transformers numbers transfer. Keep speech and non-speech audio in separate evaluation sets. ## Reader takeaway LiquidCodec is a speech-focused, streaming audio tokenizer with a useful combination of parallelizable tokens and company-reported speed and quality results. Its published model card makes it testable, but the right conclusion is not that it wins every audio workload: validate the codec against your own speech, hardware and deployment path, and do not extrapolate its speech evaluations to music or environmental sound. Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json ## Continue in the graph Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`. - topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub. - topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub. - topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub. - collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis. - collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites. - whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)