Skip to content
Audio

LFM3-Audio-2B-Full-Duplex

real-time full-duplex voice AI

3 min readupdated

An adaptation of the base LiquidAI/LFM3-Audio-2B audio model by MAdel121 that turns it into a full-duplex voice conversationalist with an 80 ms processing step. Unlike the original model, which strictly takes turns, this version listens and speaks simultaneously: it yields when interrupted, doesn't get derailed by brief user acknowledgments, and drops in natural "uh-huh" and "yeah" cues on its own without external VAD (voice activity detection), ASR (speech recognition), or TTS (text-to-speech) modules. It relies on a lightweight LoRA adapter (a small add-on to the base model, 68 MB) and a separate 91K-parameter Conv1D head to generate backchannel reactions. Built for researchers and local voice assistant developers who need natural, low-latency conversation, including on Apple Silicon.

Key features

  • 80 ms streaming mode: takes in a 16 kHz user audio stream and generates 24 kHz response audio via LiquidCodec inside a single continuous conversation.
  • Resilient to backchanneling: window-based training lets the model keep speaking through brief user interjections like "yeah, right" in 93.3% of cases, compared to 58.3% with a frame-by-frame approach.
  • Isolated reactions: a separate Conv1D head plays back 3 pre-encoded "uh-huh/yeah" audio clips while bypassing the main model's KV cache, avoiding context pollution and keeping future responses on track.
  • Base knowledge preserved: self-distillation using the model's own responses with a KL anchor eliminated persona leakage and retained roughly 95% of the base model's accuracy on spoken QA (29–30 / 120 vs. 31 / 120).

What you can do

  • Run a local browser voice chat: test live conversation and barge-in behavior on headphones via the built-in eval.live web interface.
  • Process pre-recorded audio: run .wav and .flac files offline through eval.demo to evaluate turn-taking timing and interruption handling.
  • Embed in a research assistant prototype: use the out-of-the-box 80 ms runtime as a compact alternative to heavy cascades of separate ASR and TTS models.

Example

# Запуск живого голосового чата в браузере (требуются наушники)
python -m eval.live serve --out talks/local --key live

# Офлайн-тест на аудиофайле
python -m eval.demo run --champion --in user_input.wav --out demo_out
Read moreDetails
Benchmarks (Full-Duplex-Bench v1.0, 727 clips): pause TOR of 0.044 (synthetic) and 0.116 (CANDOR), interruption relevance of 4.38 / 5, and a median interruption stop time of 1.22 s.
Hardware and speed: runs locally in float32 on Apple Silicon (MPS) at 59 ms per 80 ms step; session length is capped at roughly 3 minutes due to KV cache growth.
Training and size: the LoRA (rank 16, 68 MB) was trained in 45 GPU minutes, and the Conv1D head weighs 1.4 MB; requires the LiquidAI/LFM3-Audio-2B weights and Python 3.12.
Limitations and license: non-commercial research only (LFM Open License v1.0 + non-commercial datasets); wired headphones are mandatory in the browser since echo cancellation is disabled.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 3.8 GB · ready files4.7 GB
  • FP8 / INT8recommendedweights 1.9 GB2.6 GB
  • 4-bitweights 1.1 GB1.7 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

This is a LoRA: counted together with its base model LFM3-Audio-2B.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # MAdel121/LFM3-Audio-2B-Full-Duplex
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download LiquidAI/LFM3-Audio-2B --local-dir ~/models/LFM3-Audio-2B
    hf download MAdel121/LFM3-Audio-2B-Full-Duplex --local-dir ~/models/LFM3-Audio-2B-Full-Duplex
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to LFM3-Audio-2B
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

Requires the base model

LiquidAI/LFM3-Audio-2B

This model is a LoRA / adapter. Download the base model first, then load this one on top.

New posts about versions of LFM3-Audio-2B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 7All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

MAdel121/LFM3-Audio-2B-Full-Duplex
Author
MAdel121
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Audio

Sayso

Sayso