README.md

metadata

title: LFM3-Omni-2B
emoji: 🎙️
colorFrom: purple
colorTo: indigo
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: 'Any-to-any: text/image/video/audio in, text or speech out'
python_version: '3.12'
startup_duration_timeout: 30m

🎙️ LFM3-Omni-2B — any input, text or speech out

Interactive demo for LiquidAI/LFM3-Omni-2B, Liquid AI's omni-modal 2.2B-parameter model. One single network understands text, images, video (including its soundtrack) and audio, and can answer in text, 24 kHz speech, or interleaved text and speech — all with one set of weights.

What you can try

• Voice chat — record or upload a question and get a spoken reply.
• Image QA — upload a photo and ask about it.
• Video QA (see + hear) — attach a clip; the model reads the frames and the audio track.
• ASR — transcribe a recording.
• TTS — read a sentence aloud.

Pick the Answer modality in Advanced to choose text, speech, or both. Speech output uses the authors' recommended audio_temperature=0.4 and audio_repetition_penalty=1.2.

Example assets & attribution

• assets/sign_beach.jpg, assets/chef_wok_flames.mp4 — CC0 samples from linoyts/repo-to-space-example-inputs and linoyts/repo-to-space-example-videos.
• assets/librispeech_question.wav — from LibriSpeech (via hf-internal-testing/librispeech_asr_demo), licensed CC BY 4.0; © LibriSpeech contributors.

Model license

The model weights are released under the LFM Open License v1.0 — check the model card for terms.