Liquid AI · October 27, 2026 · 2.17B · On device
LFM3 Omni: published specs and a local setup
Liquid AI's October 27, 2026 post introduces LFM3, and with it LFM3-Omni-2B: one model that reads text, images, video and audio and answers in text, speech, or both. Nano AI has not run it. The numbers below are from Liquid's blog and the Hugging Face model card.
LFM3 Omni is not loaded here
Need an image? Use Nano Banana 2.1
LiquidAI/LFM3-Omni-2B runs on your machine, under the LFM Open License v1.0. Nano Banana 2.1 makes an image in the browser for 8 credits. It does not load the omni checkpoint.
- Made by
- Liquid AI
- Released
- October 27, 2026
- Run on Nano AI
- Nano Banana 2.1 · 8 credits per image
- Page updated
- October 31, 2026
Source: liquid.ai
What LFM3 Omni publishes
The model card for LiquidAI/LFM3-Omni-2B lists 2.17B parameters, a 32K context that counts input and output together, and a 128,000-token text vocabulary. Inputs are text, images, video with or without its soundtrack, and 16 kHz audio. Video is sampled at 2 frames per second, 4 to 384 frames by default. Outputs are text, 24 kHz speech, or interleaved text and speech. The card says it is a non-reasoning model and that enable_thinking=True is not supported. The license line is LFM Open License v1.0. Read that license in the repo before you ship anything; this page does not restate its terms.
Liquid's blog the same day says the first LFM3 release is a set of models around 2B parameters on one backbone: text, vision, audio, omni and decision. It gives the text model a 256K context and says the text model fits in about 1 GB at 4-bit. Those two figures are about the text model, not about LFM3-Omni-2B's 32K card. The omni section describes spoken questions about a picture, video watched together with its soundtrack, and tool calls from what the model sees and hears. This site has not measured latency, word error rate or battery use.
The card's install line is pip install "transformers>=5.20.2" plus torch, torchvision, torchcodec, accelerate, pillow, librosa and soundfile, then AutoModelForMultimodalLM and AutoProcessor with trust_remote_code. A GGUF variant is listed for llama.cpp. Speech sampling defaults in the card are audio_temperature=0.4 and audio_repetition_penalty=1.2. None of that runs on Nano AI. If you need an image, Nano Banana 2.1 on this site costs 8 credits.
LFM3 Omni at a glance
- Checkpoint
- LiquidAI/LFM3-Omni-2B
- Announced
- October 27, 2026, with the LFM3 family
- Parameters
- 2.17B, per the model card
- Context
- 32K tokens, input plus output
- Audio
- 16 kHz in, 24 kHz speech out
- Video
- 2 fps, 4 to 384 frames by default
- License
- LFM Open License v1.0
- On Nano AI
- Not hosted. Images use Nano Banana 2.1 at 8 credits
Load LFM3-Omni-2B the way the card shows
Copied from the Hugging Face model card as read on October 31, 2026. Nano AI does not execute it.
- 1
Install the Python packages
The card's install line, including transformers 5.20.2 or newer.
pip install "transformers>=5.20.2" torch torchvision torchcodec accelerate pillow librosa soundfile - 2
Load the checkpoint
The card's constructors. trust_remote_code is required by that snippet. device and dtype are defined just above it in the card.
model_id = "LiquidAI/LFM3-Omni-2B" model = AutoModelForMultimodalLM.from_pretrained( model_id, dtype=dtype, device_map=device, trust_remote_code=True ).eval() processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True) - 3
Speech defaults
The card's generation_config defaults for speech. Text output can stay greedy.
audio_temperature=0.4 audio_repetition_penalty=1.2 - 4
GGUF is a different file
The card lists a GGUF build for llama.cpp, including a realtime speech server. Use the files Liquid published in the same repo, not a renamed copy.
Variants the card names
Both rows link to the model repo. File sizes are not copied here because the card's table does not print a byte size.
What this review does not claim
No score from this site
Liquid's blog describes the model. Nano AI has not timed it, has not scored speech, and has not compared it with a hosted API on the same prompts.
Do not borrow the text model's context
256K and the about-1-GB 4-bit note in the blog are for the LFM3 text model. The omni card says 32K.
Thinking mode is off
The card says this checkpoint is non-reasoning and rejects enable_thinking=True.
Images here are Nano Banana
The button on this page opens Nano Banana 2.1 at 8 credits. It does not load LiquidAI/LFM3-Omni-2B.
Omni card versus the text-model lines
Both columns are Liquid AI's, from the October 27, 2026 blog and the LFM3-Omni-2B card, as read on October 31, 2026.
| LFM3-Omni-2B | LFM3 text model, same blog | |
|---|---|---|
| Parameters | 2.17B | The blog's 2B-class set |
| Context | 32K | 256K |
| Inputs | Text, image, video, audio | Text |
| Speech out | 24 kHz | Not the omni speech path |
| License line | LFM Open License v1.0 | Open weights, per the blog |
How to keep going if you wanted a picture
- 1
Run the checkpoint locally if you need omni
Use the install block below and Liquid's model card. Nano AI will not download the weights for you.
- 2
Or open Nano Banana 2.1
That model costs 8 credits an image. It takes a text prompt. It does not hear audio or watch a video file.
- 3
Read the license
The card names LFM Open License v1.0. The license file is the terms, not this paragraph.
Image prompts for Nano Banana 2.1
Only for the image button. They are not LFM3 inputs and they contain no audio.
Radio on a sill
A small wooden radio on a sunlit windowsill, linen curtain, soft dust, no people, no readable text, no logos.
Trail map
A folded blank trail map and a compass on a wooden bench, overcast light, no place names, no people, no logos.
Headphones
Over-ear headphones on a closed notebook, top-down, grey desk, soft shadow, no logos, no readable text, no people.
Kitchen speaker
A compact speaker on a kitchen shelf beside a plant, morning light, no brand marks, no people, no readable text.
If the checkpoint is not what the blog sentence suggested
- Context stops near 32K
- That matches the omni card. The 256K figure is the text model in the same blog.
- enable_thinking fails
- The card says the flag is unsupported on this non-reasoning checkpoint.
- You wanted the model inside Nano AI
- It is not loaded. The image button is Nano Banana 2.1 at 8 credits.
LFM3 Omni is not loaded here
Need an image? Use Nano Banana 2.1
LiquidAI/LFM3-Omni-2B runs on your machine, under the LFM Open License v1.0. Nano Banana 2.1 makes an image in the browser for 8 credits. It does not load the omni checkpoint.
FAQ
LFM3 Omni FAQ
Card facts only. This site has not run the model.
What checkpoint is LFM3 Omni?
LiquidAI/LFM3-Omni-2B. The card says 2.17B parameters and a 32K context. Liquid announced the LFM3 family on October 27, 2026.
Does Nano AI host it?
No. The button on this page opens Nano Banana 2.1 at 8 credits an image. Eligible new accounts get 64 credits, enough for 8 images. That is not an LFM3 session.
What are the audio rates?
The card says 16 kHz audio in and 24 kHz speech out. It also lists audio_temperature=0.4 and audio_repetition_penalty=1.2 for speech.
Is the 256K context the omni model?
No. The October 27, 2026 blog attaches 256K to the text model. The omni model card says 32K.
Can it think step by step?
The card says it is non-reasoning and that enable_thinking=True is not supported. This page has not tried to override that.
Is Nano AI affiliated with Liquid AI?
No. Nano AI is not affiliated with Liquid AI. Specs are quoted from their blog and model card as read on October 31, 2026, and this site has not benchmarked the checkpoint.
More from Nano AI
Related pages
Nano AI is an independent service and is not affiliated with Liquid AI. LFM3 Omni figures, including 2.17B, 32K, 16 kHz, 24 kHz and LFM Open License v1.0, are from Liquid's October 27, 2026 blog and the LFM3-Omni-2B model card as read on October 31, 2026. This site has not run the checkpoint.