Liquid AI LFM3: Tiny 2B models for speech, vision and video land in Ollama

What this is about

On October 27, 2026, three new models from Liquid AI's LFM3 family appeared in the Ollama library: lfm3-audio, lfm3-omni and lfm3-vl. At around 2 billion parameters each, all three are very small. Parameters are, roughly speaking, the dials a model tuned during training. They're meant to run directly on your device, meaning laptops, phones and other edge hardware (edge means the computation happens locally, not in the cloud). According to the model pages, the LFM3 family also includes Base, Text and Decision models, but those aren't covered here.

Start from the command line: ollama run lfm3-audio, ollama run lfm3-omni or ollama run lfm3-vl. Download counts were still minimal when this was written (11 to 24), and the models had been published just over 30 minutes earlier.

LFM3-Audio-2B: speak, listen, translate

LFM3-Audio-2B is the main model in this story. It takes text or audio and replies with text, spoken language, or a mix of both. According to its description, it can do the following:

• Voice chat: You ask a question out loud and get the answer streamed as text and speech. For text-only replies, the model can "think" before answering.
• Multilingual speech processing: Transcribe speech (convert it to text), synthesize it (generate it from text) and translate it, in 18 languages. The list includes German, French, Spanish, Italian, Japanese, Korean, Chinese, Arabic and more. English is listed in British and American variants, so the table counts 19 entries for 18 languages.
• Audio understanding: Interpret recordings and answer questions about speech, sounds and events in them.
• Tool calling: Spoken requests are turned into tool calls to control apps or devices, accompanied by text or speech.

New compared with its predecessor LFM2.5-Audio are an in-house audio encoder (the part that turns sound into a form the model can process) and the audio tokenizer "LiquidCodec" (it breaks speech into discrete building blocks called tokens). This is meant to improve speech quality. That's a vendor claim, and the page offers no measurements.

Key specs

Context is the amount of text the model can keep in view at once (input and output combined). The audio model works in streaming mode: it encodes audio causally (using only what it has already heard), fills the context step by step and generates speech continuously. That lets replies start before the input has been fully processed.

LFM3-Omni-2B: everything in one model

According to Liquid AI, Omni is the first omni-modal model in the LFM3 family. It understands text, images, video and audio, and replies in text and speech. The pitch: one model replaces the usual chain of ASR (speech recognition), a vision-language model, a language model and TTS (text-to-speech). With videos that have sound, it sees and hears at the same time, so it can answer questions that need both. Images and video are processed at native resolution; each 2×2 block of image patches becomes one language token, and video is sampled at 2 frames per second. The text vocabulary has 128,000 tokens. The model understands text in 16 languages, including German; speech output covers the same languages as the audio model. Thinking is explicitly not supported.

LFM3-VL-2B: image understanding with a thinking mode

The vision-language model combines the LFM3 language model with a new in-house vision encoder. According to the page, the new features are:

• Video understanding: marked "soon," so not yet available. It would be the first LFM vision model with video input.
• Thinking: The model reasons step by step before answering, which should help with math, charts and multistep tasks.
• Spatial reasoning: Finding objects and judging their positions relative to each other, intended as a foundation for robotics.

Architecturally, the model has 24 layers: 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. That mix is meant to be less compute-hungry than pure attention models. The context is 128K tokens. Use cases according to the maker: reading documents, charts and app screens, finding objects in images, and spatial reasoning for robots. Not recommended for coding, multi-turn conversations or complex reasoning.

What you need to try it

The source gives no VRAM (video memory) figures. At around 2 billion parameters, a rough estimate with 4-bit quantization is about 3 GB of video or system memory, more with longer context. That should work on a normal laptop without a dedicated GPU, on an Apple Silicon Mac or on CPU alone, though speed depends on your hardware. Plan for a few GB of free disk space. For voice chat, you need a microphone and speakers, or an interface that connects audio input and output to Ollama. The page doesn't say how far Ollama itself supports speech output.

Our take

A 2B model that combines speech input, speech output and, in Omni's case, images and video in one network is interesting for local assistants. It saves you the chain of several models, and with it memory and latency. German is supported for both text and speech. How good the quality really is remains open: the Ollama pages contain no benchmarks, comparisons with competitors or audio samples, so all quality claims are vendor promises. The maker itself warns that the models rely on search or retrieval for knowledge-intensive tasks and, because of their size, aren't suited to agentic coding or complex multistep tasks. The LFM Open License v1.0 is a custom license, so read its terms before any commercial use. Also, VL's video feature isn't available yet.