HeyHomees Lab Notes · No. 15 · Liquid AI · Published November 3, 2026
By HeyHomees Newsroom ·
In 30 seconds
- 2.17B-parameter omni-modal model runs on phones and laptops.
- Processes text, images, video (2 fps) and 16 kHz audio, outputs 24 kHz speech.
- Non-reasoning model; generation modes selectable via system prompts.
What is it?
LFM3-Omni-2B is Liquid AI's first omni-modal model that understands text, images, video and audio and answers in text and speech, replacing separate ASR, vision-language, LLM and TTS models.
How does it work?
The model uses an LFM3 backbone with a vision encoder, causal audio encoder and LiquidCodec audio tokenizer, accepting inputs at native resolution and sampling video at 2 fps; each 2×2 patch block becomes one language token.
What does it need to run?
Designed for on-device use, it runs efficiently on phones, laptops and other edge hardware; inference can use BF16 weights with SGLang on GPU or GGUF with llama.cpp on CPUs/GPUs, including a realtime speech server.
What's new versus the last version?
It is the first omni-modal model in the LFM3 family, combining text, vision, audio, and decision capabilities in a single 2.17B-parameter checkpoint, whereas prior family members were single-modality (base, text, vision, audio) models.
By the numbers
- Parameters: 2.17B total
- Context length: 32K tokens (input + output)
- Vocabulary size: 128,000 text tokens
- Speech output sample rate: 24 kHz speech
- Video sampling rate: 2 fps
The honest read
The release provides only architectural and capability specifications; it does not include benchmark scores, so performance claims are limited to the model’s size, context length, and supported modalities as self-reported by Liquid AI. The model is described as non-reasoning, meaning it does not perform chain-of-thought or external tool use, and all generation settings are based on user-provided prompts rather than internal evaluation metrics.
What we don't know yet
- No quantitative accuracy or latency benchmarks are provided.
- Training data composition and amount are not disclosed.
- Exact download size of the checkpoint is not stated.
Sources
- Model card, Hugging Face (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.