LFM3-Omni-2B

Use it to transcribe or translate speech, discuss images by voice, or ask questions about a video's visuals and soundtrack.

Availability

Checkpoint available on Hugging Face; API and playground links provided

Pricing model

Not established

Deployment

Designed for on-device use; local inference via llama.cpp, and API access is linked

Privacy

Not established

Licence

LFM Open License v1.0

• The model accepts text, images, video with or without its soundtrack, and 16 kHz audio. ^[1]

  <td style="border: 0; padding: 0.55em 0.75em;">Text, images, video with or without its soundtrack, and 16 kHz audio.

• Outputs include text, 24 kHz speech, or interleaved text and speech. ^[1]

  <td style="border: 0; padding: 0.55em 0.75em;">Text, 24 kHz speech, or interleaved text and speech</td>

• The model can process video visuals and audio together to answer questions requiring both. ^{[1] [2]}

  - **Watches and listens together:** for videos with sound, the model sees and hears them at the same time, so it can answer questions that need both.

• BF16 and GGUF checkpoints are provided for inference. ^[1]
• GGUF supports local inference on CPUs and GPUs, including a realtime speech server. ^[1]

  <td style="border: 0; padding: 0.55em 0.75em;">Local inference on CPUs and GPUs, including a realtime speech server.</td>

• The model is non-reasoning; enable_thinking=True is not supported. ^[1]

  `enable_thinking=True` is not supported.

• New models bring live speech translation and voice features Daily brief 2026-10-28