LFM3 Omni: run Liquid’s picture-and-voice model locally on a Mac

What LFM3 Omni is

LFM3 Omni is Liquid AI’s model for mixed inputs: it takes pictures and audio alongside text and answers in text. It has around two billion parameters (Liquid lists it as LFM3-Omni-2B) and sits in the LFM3 family next to separate text, vision, audio and decision models. The reason to pick Omni is that one model handles those inputs together, so you can show it something and explain it out loud instead of typing everything into a prompt.

You can try LFM3 in Liquid’s Playground or through its API. This guide follows the local route from the video: you download the model files and the inputs are processed on your own Mac.

What Liquid’s benchmarks say

Liquid’s model card compares Omni with models including Qwen2.5-Omni and Gemma 4. On Daily-Omni, which tests understanding video and audio together, Omni scores about 70 against about 61 for Qwen2.5-Omni 7B. On OmniBench overall, which pairs an image with an audio clip, it scores about 46 against about 50 for the same Qwen model.

Those are Liquid’s measurements, not ours. They show a strong result for a small model on one combined test and a loss on another; they don’t tell you how it will handle your own files.

Our test: a chart and a recording

We gave Omni two sources that describe different parts of the same situation. A bar chart shows workshop sign-ups by day: Monday 12, Tuesday 24 and Thursday 18. A short recording says: “Please move the workshop to Thursday at two p.m. Maya will send the outline on Tuesday. We still have not chosen a room. Do not book one yet.” The recording is synthetic macOS speech saved as a file, not live microphone input.

The prompt asked for the most popular day, the new workshop time and whether to book a room. Omni answered: “Tuesday has the most sign-ups (24). The workshop has been moved to Thursday at 2 PM. We should not book a room yet.” It picked up the change in the recording without replacing what the chart says, which is the point of the test: the most popular day and the day actually chosen stay separate.

Run the same test on your Mac

We ran this on 2 November 2026 with llama.cpp build b11960 on an M5 Pro MacBook Pro with 48 GB of RAM. That is the machine we used, not a minimum requirement. It will probably run on your laptop too; just expect the computer to be busy for a minute or so after you press Enter.

1. Download the Mac ARM64 build of llama.cpp (release b11960) and extract it. Older runner versions may not support LFM3.
2. From Liquid’s LFM3-Omni-2B-GGUF repository on Hugging Face, download LFM3-Omni-2B-Q4_K_M.gguf (the model) and mmproj-LFM3-Omni-2B-Q8_0.gguf (the matching projector). The picture and audio route needs both files. Together they are about 1.6 GB, and running them takes additional memory.
3. Download the free starter files above and put workshop-chart.png, instruction.wav and omni-mixed-prompt.txt in the same folder as the two model files.
4. Open Terminal in that folder and run: ./llama cli -m LFM3-Omni-2B-Q4_K_M.gguf --mmproj mmproj-LFM3-Omni-2B-Q8_0.gguf -c 4096 -n 500 --reasoning off --temp 0 --seed 42 --single-turn --image workshop-chart.png --audio instruction.wav -sys 'Answer from the supplied image and audio only.' -f omni-mixed-prompt.txt Replace ./llama with the path to the llama executable you extracted.
5. The model and projector load, then the answer appears in Terminal. Compare it with the chart and the recording.
6. To try your own material, replace the image and recording in the command and change the questions in the prompt file. Start with sources you can check yourself.

The chart-scale mistake

We also asked for the total number of sign-ups from the chart alone. Omni read 12, 24 and 18 and correctly added them to 54, then said that meant 27 people. Each coloured box represents two people, but the labels already count people, so it applied the scale a second time.

When we pointed that out and asked it to double-check, it corrected the answer to 54 people. That is a correction with feedback, not the model catching its own mistake. Keep the source image beside the answer, especially for figures.

Small model, bigger models and what to try next

Simple mistakes like this are more likely from a model this small. Bigger downloaded models usually need a better computer: a model of about 2 GB can run on most computers, while Luke has run models of up to about 30 GB on this MacBook and finds them more reliable. That is his experience, not a benchmark.

Omni’s model card also lists video input and speech output through compatible runners. We didn’t test either; comment on the video if you’d like to see them covered.

This guide is free, with no email gate. Follow the steps yourself, or paste the guide into ChatGPT, Codex, Claude or another assistant and ask it to help you set it up. Read each command before you run it.

The takeaway

Omni can combine what it sees in a picture with what it hears in a recording on a local Mac, and our combined question came back right. Its chart arithmetic slipped once, so check figures against the source and give feedback when something doesn’t add up.