Liquid AI releases LFM3-Omni-2B, a 2.17B omni-modal model with speech output

Liquid AI has released LFM3-Omni-2B, a post-trained omni-modal model with 2.17B total parameters and a 32K-token context window (input plus output). It is the first omni-modal model in the LFM3 family and takes text, images, video and audio as input. It can answer in text, 24 kHz speech, or interleaved text and speech.

Liquid AI claims the single end-to-end model replaces a separate ASR model, vision-language model, LLM and TTS model, and says it is designed to run efficiently on phones, laptops and other edge hardware.

Specifications

• Parameters: 2.17B total
• Architecture: LFM3 backbone, LFM3 vision encoder, LFM3 causal audio encoder, and the LiquidCodec audio tokenizer
• Context length: 32K tokens (input plus output)
• Vocabulary: 128,000 text tokens
• Inputs: text, images, video with or without its soundtrack, and 16 kHz audio
• Outputs: text, 24 kHz speech, or interleaved text and speech
• License: LFM Open License v1.0
• Reasoning: none. This is a non-reasoning model, and enable_thinking=True is not supported.

Images and video are processed at native resolution, and each 2×2 block of patches becomes one language token. Video is sampled at 2 fps, with 4 to 384 frames by default. For videos with sound, the model processes the visuals and audio together, which Liquid AI says lets it answer questions that depend on both.

Language support

Text languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian and Polish.

Speech languages and locales: Arabic, Danish, German, Greek, British English, American English, Finnish, French, Hebrew, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Spanish, Swedish, Turkish and Chinese.

Use cases and prompting

The model card lists voice questions about a photo, speech transcription and translation, spoken chat, and questions about a video and its soundtrack. ASR and TTS are driven by system prompts: Perform ASR in <language>. and Perform TTS in <language> using a <gender> voice. Output modality is chosen by default from the system prompt and input modality, typically matching the input. An optional generation_mode argument forces text, audio or interleaved output.

For text output, Liquid AI says greedy decoding works well. For speech or interleaved output, it recommends audio_temperature=0.4 and audio_repetition_penalty=1.2.

Checkpoints and inference

Two checkpoint variants are available:

• BF16: weights for GPU inference with SGLang or Transformers. SGLang serves text output.
• GGUF: for llama.cpp, with local inference on CPUs and GPUs, including a realtime speech server.

The Transformers path requires transformers>=5.20.2 and trust_remote_code=True. The codec implementation ships in the model repository, and the processor loads its weights from audio_tokenizer/.

The LFM3 family includes base, text, vision, audio, omni and decision models. Benchmark scores and API pricing are not published on the model page.

What this means

LFM3-Omni-2B consolidates speech recognition, vision-language understanding, text generation and speech synthesis into one 2.17B-parameter checkpoint. For teams building voice or multimodal features for local hardware, that means one set of weights and one runtime instead of a chained pipeline of separate models.

The GGUF build with a realtime speech server and the 16 kHz input and 24 kHz output audio specs make local voice interaction a documented use case. Because the model is non-reasoning, it targets perception and conversation tasks rather than extended chain-of-thought work. No benchmark results are published on the model page, so claims about quality against separate specialist models remain Liquid AI's own until independent evaluations appear.