---
type: "post"
id: "ships-q4-0-q4-k-m-q8-0-c575d2"
slug: "ships-q4-0-q4-k-m-q8-0-c575d2"
title: "Ships Q4_0, Q4_K_M, Q8_0 GGUF builds for llama.cpp on-device inference."
description: "Liquid AI introduced LiquidCodec, an approximately 74-million-parameter audio codec designed for its speech models and on-device inference."
retrieval_nugget: "Liquid AI introduced LiquidCodec, an approximately 74-million-parameter audio codec designed for its speech models and on-device inference. The company reports fast CPU encoding and decoding, strong reconstruction and text-to-speech results in its evaluations, and quantized GGUF builds for llama.cpp—useful signals for teams weighing local voice generation against hosted services."
published_at: "2026-11-13"
updated_at: "2026-11-13"
record_date: "2026-11-13"
date_kind: "published_at"
topics: ["inference"]
entities: ["Liquid AI","LiquidCodec","LFM3-Audio-2B","LFM3-Omni-2B","llama.cpp","Qwen3-1.7B"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "Liquid AI released LiquidCodec GGUF builds in Q4_0, Q4_K_M, and Q8_0 formats for on-device inference with llama.cpp."
prior_refs: []
source_url: "https://www.liquid.ai/blog/liquidcodec"
source_urls: ["https://www.liquid.ai/blog/liquidcodec"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:ships-q4-0-q4-k-m-q8-0-c575d2"
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/ships-q4-0-q4-k-m-q8-0-c575d2/","markdown":"https://newruntime.com/posts/ships-q4-0-q4-k-m-q8-0-c575d2.md","json":"https://newruntime.com/posts/ships-q4-0-q4-k-m-q8-0-c575d2.json"}
markdown_url: "https://newruntime.com/posts/ships-q4-0-q4-k-m-q8-0-c575d2.md"
---

# Ships Q4_0, Q4_K_M, Q8_0 GGUF builds for llama.cpp on-device inference.

## Retrieval answer

Liquid AI introduced LiquidCodec, an approximately 74-million-parameter audio codec designed for its speech models and on-device inference. The company reports fast CPU encoding and decoding, strong reconstruction and text-to-speech results in its evaluations, and quantized GGUF builds for llama.cpp—useful signals for teams weighing local voice generation against hosted services.

## Summary
Liquid AI introduced LiquidCodec, an approximately 74-million-parameter audio codec designed for its speech models and on-device inference. The company reports fast CPU encoding and decoding, strong reconstruction and text-to-speech results in its evaluations, and quantized GGUF builds for llama.cpp—useful signals for teams weighing local voice generation against hosted services.

## What happened
Liquid AI says it built LiquidCodec for its latest audio models, LFM3-Audio-2B and LFM3-Omni-2B, which it released at the end of October. An audio codec converts speech into tokens that a language model can predict, then reconstructs audio from those tokens. The company describes LiquidCodec as an approximately 74-million-parameter model, based on its LFM3 architecture, with an encoder and decoder designed to be symmetric and the encoder fully causal. It processes 24 kHz waveforms at 12.5 frames per second, or one frame every 80 milliseconds.

The design combines spectral processing with grouped finite scalar quantization (GFSQ). Liquid AI contrasts this with residual vector quantization (RVQ), which it says introduces stronger sequential dependencies between codebooks. GFSQ splits feature vectors into independently quantized groups, allowing parallel prediction of audio tokens. The codec’s decoder predicts spectral coefficients directly, which the company says avoids costly progressive upsampling operations. Liquid AI says it previously replaced only the detokenizer in LFM2.5-Audio-1.5B and achieved eight-times-faster detokenization; LiquidCodec replaces both tokenizer and detokenizer for the newer generation.

The company trained the codec in two stages. The first used 264,000 hours of material across 107 languages, including web speech, audiobooks, crowd-sourced recordings, and smaller parliamentary and meeting collections. The second fine-tuned on 55,000 hours of cleaner, full-band speech sampled natively at 24 kHz or above, using adversarial critics focused on waveform patterns, spectral detail, and speech features. Liquid AI also used decoder-only quantization-aware training to improve robustness for edge inference.

For speed measurements, Liquid AI used single-threaded CPU inference in FP32 on an Apple M4 Max MacBook Pro. It reports encoding at about 361 times real time and decoding at about 87 times real time. Its reconstruction and zero-shot TTS comparisons covered LibriSpeech test-clean, Seed-TTS English and Mandarin, and FLEURS in German, French, Mandarin, Japanese, and Korean. The company says LiquidCodec achieved the best ViSQOL and mel-distance results on all eight test sets, while remaining competitive on other reconstruction metrics. TTS models trained on LiquidCodec tokens achieved the highest UTMOSv2 naturalness scores on all eight sets and best intelligibility scores on four; a repeat using a Qwen3-1.7B backbone gave consistent results.

LiquidCodec is available on Hugging Face and loads with Transformers. For on-device use, Liquid AI points to GGUF builds of LFM3-Audio-2B and LFM3-Omni-2B with Q4_0, Q4_K_M, and Q8_0 quantizations for llama.cpp. The company says Q4_0 can reduce the decoder footprint by a factor of 3.5 with minimal impact on reconstruction quality.

## What the archive adds
New Runtime's earlier coverage includes Unsloth's Dynamic 3.0 GGUF quantizations for local Desktop and llama.cpp runs ([October 27, 2026](https://newruntime.com/posts/unsloth-ships-dynamic-3-0-gguf-quants-of-1e5a1f/)) and DiffuSpace's official Clover GGUF quantizations for llama.cpp ([October 31, 2026](https://newruntime.com/posts/diffuspace-publishes-official-clover-35b-a3b-preview-gguf-6f368d/)). LiquidCodec is a new fact in the broader local-inference thread: here the release concerns an audio codec and speech-model builds, with reported codec-specific speed and quality results.

## Trend signal
This supports the pattern that generative media is becoming application infrastructure: Liquid AI presents a codec as a component inside speech models, rather than a standalone one-off audio generator. Its performance and quantization claims matter to that direction, though they remain company-reported. [Generative media becomes application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/).

## Why it matters
A local voice feature depends on more than the language model. The tokenizer determines what speech representation the model learns to predict, while the detokenizer affects latency, compute, and memory at generation time. Liquid AI’s design makes those parts a coordinated component of its audio-model stack and offers packaged quantized model builds for a common local runtime.

For product teams, the useful implication is not that a reported benchmark guarantees a better user experience. It is that a codec can be evaluated as a deployment choice alongside the speech model: reconstruction quality, generated speech naturalness, intelligibility, decoding speed, and memory footprint all matter. The published numbers suggest a candidate worth testing for low-latency local speech, especially where connectivity or cloud cost is a constraint. Teams should validate the codec and complete model on their own hardware, languages, and audio conditions before committing.

## Limits and uncertainty
The benchmark results and speed figures are reported by Liquid AI; they are not independently verified here. The speed test specifies an M4 Max, single-threaded CPU inference, and FP32, so it does not establish performance across other processors or quantized deployments. The announcement does not provide a general end-to-end latency or memory comparison for complete applications, nor does it establish how the models perform on every language, microphone, or noisy environment. The phrase “minimal impact” for Q4_0 is not quantified as a single universal threshold; teams should inspect the per-dataset results and run their own tests.

## Skill to build now
Build a small codec evaluation harness before choosing a local speech stack. Run a representative set of recordings through encode-and-decode reconstruction, then measure real-time factor, memory use, and perceived artifacts on the target device. If testing generated speech, keep the text, model, prompts, and generation settings fixed across codecs; compare intelligibility as well as naturalness rather than relying on one aggregate score.

## Reader takeaway
LiquidCodec is a concrete audio infrastructure release: Liquid AI pairs its own tokenizer and detokenizer with quantized llama.cpp builds for local speech models. Its reported results make it worth evaluating, not an automatic choice. Measure full-model latency, memory, and speech quality on the hardware and languages your product actually needs.

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: [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)