llama.cpp pre-release builds now package support for Liquid AI’s new LFM3-Audio-2B and LFM3-Omni-2B models.[4][5] After downloading the files, HighVRAM’s tested speech-recognition and speech-generation commands processed audio locally on a modest home computer. A later pre-release also documents streaming voice conversations through llama.cpp’s server.[14]
The small files are as important as the feature list. Liquid AI’s Q4_K_M audio model is 1,099,626,016 bytes, and its required Q8 audio projector is 337,862,080 bytes.[7][8] Together they are 1.34GiB on disk, calculated by adding those two published file sizes and converting bytes to GiB.[8]
What runs today
LFM3-Audio-2B accepts text or audio and can return text, speech or both.[2] Liquid AI also documents speech recognition, synthesis, translation across 18 languages and spoken tool calls, although it warns that a 2B model is not meant for agentic coding or difficult multi-step reasoning.[2]
LFM3-Omni-2B adds images and video to those inputs.[3] It has 2.17 billion parameters and a 32K-token context window, and it can answer in text or 24kHz speech.[3] Liquid AI designed it to combine jobs often split among a speech recognizer, a vision-language model and a speech synthesizer. HighVRAM did not test the Omni model or image input.
There is still some assembly involved. Release b11941 contains the basic model support and packaged binaries for macOS, Windows and several Linux backends.[5] Release b11946 followed with a `/v1/realtime` server endpoint for streaming audio input, audio output and live transcription using both LFM3 variants.[13][14] HighVRAM did not test that streaming endpoint. The b11941 tools also labeled audio input and the multimodal command-line program experimental during our test, so this is early runtime plumbing rather than a finished voice assistant with a polished microphone button.
The Omni name also gets ahead of the current llama.cpp path in one place. Its GGUF documentation provides image and audio commands but says video input is not supported yet.[9] Buyers should read “omni” as the model’s capability, not a promise that every runtime exposes every modality on day one.
HighVRAM’s small test found a real limit
HighVRAM downloaded the exact b11941 Apple-silicon binary, the Q4_K_M audio model and its Q8 projector. Both model-file SHA-256 hashes matched the hashes published by Hugging Face.[8] On a 16GB M4 Mac mini, the first transcription command completed successfully and reached a maximum resident set size of 3.06GiB.
We fed it a clean, computer-generated WAV saying, “High VRAM keeps artificial intelligence at home.” The model returned, “Hyviram keeps artificial intelligence at home.” That is one synthetic sentence, not a benchmark, but it is a useful warning: a tiny local speech model can run comfortably while still mangling an unfamiliar name.
The speech-generation command also produced a valid 2.88-second, 24kHz WAV from the sentence “Local AI should stay under your control.” llama.cpp reported 1.13 seconds for prompt evaluation, generation and vocoding after the model was loaded. We verified the file format and duration; we did not judge whether the speech was intelligible, natural or competitive with a cloud service.
Those checks add two things the launch material cannot. The published package really runs, and the tested one-shot Audio command’s memory footprint makes an 8GB Apple-silicon machine plausible. Our test used 16GB, however, so 8GB remains a fit estimate rather than a measured recommendation. It does not cover Omni, the realtime server, long contexts or a computer busy with other applications. On Windows or Linux with a discrete GPU, backend choice and offload settings will change both memory placement and speed.
The Omni download is slightly larger. Its Q4_K_M weights plus Q8 projector total 1.53GiB, calculated from published sizes of 1,099,630,080 and 540,892,768 bytes.[10] Allow several gigabytes beyond either download for model mappings, runtime state, the operating system and other applications. Storage size is not the same thing as required RAM.
The benchmark and license catches
Liquid AI reports strong audio-visual benchmark results for Omni, but the company measured the published comparisons using each model’s own prompt template and preprocessing.[3] Those are vendor benchmarks. The Audio card also publishes company measurements, not independent tests.[2] They show that Liquid AI evaluated its models, but not whether either one will recognize your microphone, accent, room noise or specialist vocabulary better than Whisper, Parakeet or a cloud transcription service.
The license also needs a closer look than the phrase “open-weight” suggests. LFM Open License v1.0 permits home and research use, modification and redistribution under stated conditions.[12] Its commercial-use section defines a $10 million annual-revenue threshold, then says entities that exceed the threshold are not licensed under the agreement.[12] Commercial users approaching that figure should read the agreement carefully or obtain separate terms. Most home users will never face that issue.
Privacy is the cleaner win. Once the binaries and weights are downloaded, the transcription and speech-generation paths we tested run on the user’s machine. That makes LFM3 attractive for voice notes, home automation, accessibility experiments and private recordings that should not leave the house. Applications built around it still need sensible defaults: binding a server to the local network or adding a cloud tool can give the data a new exit.
For now, LFM3-Audio is the better first download. It is smaller, its speech job is clear, and llama.cpp exposes the capabilities that matter. LFM3-Omni becomes more interesting when an application genuinely needs images and sound in one conversation; video users should wait for the runtime gap to close. Independent word-error testing on ordinary microphones, noisy rooms and the names people actually say will decide how useful this small model is.
Sources
- [2] LFM3-Audio-2B model card
- [3] LFM3-Omni-2B model card
- [4] llama.cpp LFM3 audio and omni support PR
- [5] llama.cpp b11941 release
- [7] LFM3-Audio-2B GGUF
- [8] LFM3-Audio-2B GGUF file tree
- [9] LFM3-Omni-2B GGUF
- [10] LFM3-Omni-2B GGUF file tree
- [12] LFM Open License v1.0
- [13] llama.cpp b11946 release
- [14] llama.cpp realtime voice PR