LiquidCodec (~74M On-Device Streaming Audio Codec) Deep Dive: 24kHz / 12.5Hz / 12 GFSQ Architecture, LFM3-Audio-2B Vocoder Boundaries, and LFM v1.0 License ($10M Cap)
Examine Liquid AI's ~74M on-device streaming audio codec LiquidCodec (24kHz, 12.5Hz frame rate, 12-group GFSQ), its tokenizer/vocoder role inside LFM3-Audio-2B, 361x encoding and 87x decoding CPU speedups, Korean FLEURS benchmarks, and the $10M annual revenue commercial threshold under the LFM Open License v1.0.
LiquidCodec (~74M) Symmetric STFT · 12-GFSQ · iSTFT Pipeline, LFM3-Audio-2B Boundaries & License Guardrails Diagram (Conceptual SVG Illustration)
On November 12–13, 2026, Liquid AI published the technical architecture and evaluation benchmarks for LiquidCodec, the ~74M-parameter streaming neural audio codec that powers speech tokenization and waveform reconstruction in LFM3-Audio-2B and LFM3-Omni-2B. Operating on 24 kHz audio at a 12.5 Hz frame rate (one frame every 80 ms) with 12-group Grouped Finite Scalar Quantization (GFSQ), LiquidCodec encodes audio at approximately 361× real time (RTF 0.0028) and decodes tokens back into speech at 87× real time (RTF 0.0115) on a single Apple M4 Max CPU thread in FP32.
The essential architectural and licensing takeaway is three-fold: first, LiquidCodec (~74M) is an audio tokenizer and detokenizer (vocoder) layer—not a standalone end-to-end TTS, ASR, or multimodal language model. Second, even inside LFM3-Audio-2B, input audio understanding/ASR (16 kHz via a causal audio encoder) is architecturally separate from output speech synthesis (24 kHz via LiquidCodec in audio_tokenizer/). Third, the weights ship under the LFM Open License v1.0—not Apache 2.0—which restricts free commercial use to organizations whose corporate group (Legal Entity, including parent and common-control affiliates) has under $10,000,000 USD in annual revenue. This post synthesizes the LiquidCodec engineering blog, the LFM3 announcement, the LFM3-Audio-2B Hugging Face model card, and the LFM Open License v1.0 specification, paired with a 6-check local Python verification script rather than subjective human listening tests on proprietary voice datasets.
1. Four Boundaries to Keep Separate: Codec vs. TTS Backbone, 16kHz Input vs. 24kHz Output, GFSQ vs. RVQ, and Licensing
When evaluating LiquidCodec for voice pipelines or internal technical briefings, separate these four boundaries:
- LiquidCodec (the ~74M audio tokenizer/vocoder layer) is separate from the TTS/multimodal LLM backbone (
LFM3-Audio-2B): LiquidCodec does not accept text prompts or perform speech recognition. Its tokenizer converts24 kHzwaveforms into discrete tokens (12.5 Hz, 12 tokens per80 msframe), and its detokenizer converts predicted discrete tokens back into24 kHzspeech waveforms. Full Text-to-Speech (TTS) requires a language model backbone (such asLFM3-1.7BorQwen3-1.7Bpaired with a 6-layer, 1,024-dim autoregressive depth decoder) to predict the 12 codebook tokens per frame before callingcodec.decode()(LiquidCodec Architecture).
LFM3-Audio-2B's16 kHzcausal input encoder is separate from its24 kHzLiquidCodec output path: For incoming speech (ASR, LiveASR in0.96 schunks, and audio understanding),LFM3-Audio-2Bfeeds16 kHzaudio into a dedicated causal audio encoder built from sliding-window attention and short convolutions. For outgoing speech synthesis and interleaved voice chat, the model predicts discrete audio codes thatprocessor.audio_tokenizer(LiquidCodec) decodes into24 kHzaudio (LFM3-Audio-2B Model Details).
- Grouped Finite Scalar Quantization (GFSQ) is separate from Residual Vector Quantization (RVQ): Previous LFM audio models (
LFM2-Audio-1.5BandLFM2.5-Audio-1.5B) relied on the third-party Mimi tokenizer, which uses RVQ—where each codebook quantizes the residual error left by preceding codebooks, creating strict sequential dependencies across codebooks. LiquidCodec replaces RVQ with 12-group GFSQ, splitting the 256-dim latent vector into 12 independent groups that round each dimension to fixed scalar levels and concatenate the results. Removing cross-codebook residual chains enables iterative parallel prediction (MaskGIT-style) of multiple codebooks per frame (Why we built our own audio codec).
- Open-weight availability is separate from unrestricted commercial open-source licensing: Both
LFM3-Audio-2Band its bundledaudio_tokenizer/(LiquidCodec) are governed by the LFM Open License v1.0. Under Section 5 (Commercial Use Limitation), free commercial use applies only if your Legal Entity—defined in Section 1 to include your organization plus any parent, subsidiary, or affiliate under common control—has under $10,000,000 USD ($10M) in annual revenue. Crossing the $10M threshold requires a commercial license from Liquid AI (LFM Open License User Guide).
| Dimension | LiquidCodec (~74M) Standalone | LFM3-Audio-2B (2B Full Model) | Engineering & Legal Takeaway |
| Architectural Role | Audio Tokenizer & Detokenizer (Vocoder) (Waveform ↔ 12.5 Hz discrete tokens) | Any-to-Any Multimodal Foundation Model (ASR, TTS, voice chat, audio QA, tool calling) | LiquidCodec cannot synthesize speech from text without an LLM/TTS backbone |
| Sampling Rates | 24 kHz waveform → 12.5 Hz (12 GFSQ groups) → 24 kHz waveform | Input causal audio encoder: 16 kHzOutput speech synthesis (LiquidCodec): 24 kHz | Configure microphone/ASR input streams at 16 kHz and TTS output playback at 24 kHz |
| Quantization & Runtimes | processor.audio_tokenizer (Transformers); decoder-only QAT in Q4_0 (3.5× smaller) | BF16 (vLLM, SGLang, Transformers), GGUF ( llama.cpp Q4_0/Q4_K_M/Q8_0), MLX, ONNX | Decoder QAT Q4_0 slashes edge memory footprint by 3.5× with negligible ViSQOL drop |
| License Terms | LFM Open License v1.0 (shipped under audio_tokenizer/ in repo) | LFM Open License v1.0 (32K context, 18 languages including Korean) | Requires a paid commercial agreement if corporate group revenue is >= $10M USD/year |
2. Inside the Symmetric STFT · 8-Block LFM3 · 12-Group GFSQ · iSTFT Pipeline
LiquidCodec achieves 361× real-time encoding and 87× real-time decoding on a single CPU thread by predicting spectral coefficients directly via STFT/iSTFT—avoiding expensive progressive transposed-convolution upsampling in the time domain—and using an exactly symmetric, fully causal LFM3 block stack (LiquidCodec Architecture):
- STFT Frontend &
75 Hz → 12.5 HzDownsampling: - A
24 kHzwaveform is transformed via STFT into75 Hzspectral frames (1923channels), then passed through three convolutional layers (Conv 1923 → 256, SiLU→Conv 256 → 256, SiLU→Conv 256 → 512) that downsample the frame rate by to12.5 Hz(one frame every80 ms).
- Symmetric, Fully Causal 8-Block LFM3 Encoder and Decoder:
- Both the encoder and decoder use the exact same 8-block LFM3 pattern:
[1: ShortConv, 2: ShortConv, 3: Attention, 4: ShortConv, 5: Attention, 6: ShortConv, 7: Attention, 8: ShortConv](5 ShortConv + 3 Attention blocks). - Because the encoder is fully causal (each frame only attends to or convolves over past frames), it supports low-latency live streaming without lookahead buffering.
- 12-Group GFSQ Quantization (
RMSNorm, Linear 512 → 256→FSQ 12 groups): - The encoder projects features to a 256-dim latent vector , which is split across 12 independent FSQ groups (), yielding 12 discrete tokens per
80 msframe (150 tokens/sec).
Repeat × 6Frame Expansion & Direct iSTFT Reconstruction:- The quantized latent passes through
Linear 256 → 512andRepeat × 6to expand the frame rate back from12.5 Hzto75 Hz, then traverses the 8 symmetric decoder blocks andRMSNorm, Linear 512 → 1282to predict the spectrum for iSTFT synthesis back to24 kHz.
- Two-Stage Multilingual Training (Removing Metallic Artifacts) & Decoder QAT:
- Stage 1 (264,000 hours, 107 languages): Pre-trained using only waveform and spectral reconstruction losses (no GAN losses) to learn stable representations on noisy web/audiobook data; however, pure reconstruction losses left perceptual "metallic" averaging artifacts.
- Stage 2 (55,000 hours of clean
>= 24 kHzspeech) + QAT: Fine-tuned with adversarial critics inspecting waveform patterns, spectral detail, and speech features to eliminate metallic artifacts, followed by decoder-only Quantization-Aware Training (QAT) to shrink theQ4_0decoder footprint by 3.5×.
3. Speed and Quality Benchmarks (Including FLEURS Korean)
The table below compares single-thread CPU inference speed on an Apple M4 Max (FP32 Real-Time Factor, lower is faster) alongside reconstruction and zero-shot TTS metrics on FLEURS Korean (LiquidCodec Results).
| Audio Codec | Codebooks | Encoding RTF ↓ (Real-Time Speed ↑) | Decoding RTF ↓ (Real-Time Speed ↑) | FLEURS Korean Reconstruction & Zero-Shot TTS |
| LiquidCodec (BF16) | 12 (GFSQ) | 0.0028 ± 0.0001(~361× real time) | 0.0115 ± 0.0001(~87× real time) | SQMOS 4.217 · PESQ 2.637 · Mel Dist 0.117ViSQOL 4.176 · Recon CER 4.45% · TTS UTMOSv2 2.97 |
| LiquidCodec (Q4_0 QAT) | 12 (GFSQ) | Same tokenizer | 3.5× smaller decoder | SQMOS 4.214 · PESQ 2.599 · Mel Dist 0.118ViSQOL 4.146 · Recon CER 4.30% · TTS UTMOSv2 2.88 |
| X-Codec 2 | 1 | 0.7359 ± 0.0103(~1.36× real time) | 0.0379 ± 0.0009(~26.4× real time) | SQMOS 4.214 · PESQ 2.299 · Mel Dist 0.128ViSQOL 3.870 · Recon CER 5.08% · TTS UTMOSv2 2.87 |
| NanoCodec | 13 | 0.2145 ± 0.0057(~4.66× real time) | 0.2543 ± 0.0061(~3.93× real time) | SQMOS 4.212 · PESQ 2.589 · Mel Dist 0.137ViSQOL 4.018 · Recon CER 3.41% · TTS UTMOSv2 2.86 |
| Mimi (12 codebooks) | 12 (RVQ) | 0.0361 ± 0.0004(~27.7× real time) | 0.0329 ± 0.0005(~30.4× real time) | SQMOS 3.843 · PESQ 2.165 · Mel Dist 0.178ViSQOL 3.530 · Recon CER 9.37% · TTS UTMOSv2 2.54 |
4. Practical Integration Code and Local Python Verification
4.1 Synthesizing 24 kHz Korean Speech with LFM3-Audio-2B and LiquidCodec in Transformers
When loading LiquidAI/LFM3-Audio-2B via AutoProcessor, the processor loads the LiquidCodec weights from audio_tokenizer/ and exposes the codec at processor.audio_tokenizer (LFM3-Audio-2B Transformers quick start).
4.2 Local Python Verification Script (6 Architectural & License Invariants)
The Python script below verifies six concrete invariants from the official documentation: 80 ms frame duration, 6× repeat upsampling ratio (75 Hz / 12.5 Hz), 150 tokens/s rate across 12 GFSQ groups, zero cross-group error propagation in GFSQ, single-thread RTF speedup math, Q4_0 3.5× footprint reduction and FLEURS Korean scores, 16 kHz input vs. 24 kHz output sample-rate separation, and common-control corporate revenue aggregation under the LFM Open License v1.0.
5. Technical Review & Adoption Checklist
Position LiquidCodec accurately as an audio tokenizer/vocoder layer: In architecture reviews, distinguish LiquidCodec (~74M waveform-to-token and token-to-waveform codec) from full TTS/ASR foundation models (LFM3-Audio-2B).
Separate input (16 kHz) and output (24 kHz) sample rates: Feed 16,000 Hz audio chunks into LFM3-Audio-2B's causal audio encoder for ASR/LiveASR, and write codec.decode() outputs at 24,000 Hz.
Evaluate Q4_0 QAT builds for edge deployment: Use the quantized GGUF (Q4_0, Q4_K_M, Q8_0) or QAT decoder to cut decoder memory by 3.5× while preserving ViSQOL and intelligibility.
Verify corporate group revenue against the $10M USD threshold in LFM Open License v1.0: Confirm whether your organization (including parent and common-control affiliates under the license's Legal Entity definition) exceeds $10M USD in annual revenue before using LFM3 or LiquidCodec weights in commercial products.
6. Official References (Verified 2026-11-17)
- Hugging Face Model Card — LiquidAI/LFM3-Audio-2B (
audio_tokenizer, 16kHz input / 24kHz output, TTS & LiveASR examples)
Korean version:
LiquidCodec (~74M 온디바이스 스트리밍 오디오 코덱) 아키텍처 해부: 24kHz·12.5Hz·12 GFSQ 양자화, LFM3-Audio-2B 보코더 경계 및 LFM 라이선스($10M) 점검하기