# LFM3 on Ollama: Liquid AI's four-model family that runs on anything

Liquid AI has dropped four LFM3 models on Ollama at once — text, vision, audio and omni — and every one of them is small enough for a phone. Here's what each does and what it needs.

Published 2026-10-29 · https://inpherio.co.uk/blog/lfm3-on-ollama-liquid-ai-s-four-model-family-that-runs-on-anything

Key points:
- Liquid AI's LFM3 family landed on Ollama on 29 October 2026: four models (text, vision, audio, omni), all 2.4GB or under, all under the LFM Open License v1.0.
- Every model is designed for phones, laptops and edge hardware — `ollama run lfm3` (or -vl, -audio, -omni) is all it takes.
- lfm3's 256K context and agentic post-training are the headline features; the readmes are equally clear it's not for coding or complex reasoning.
- lfm3-omni is the most interesting: one 2.17B model replacing ASR, VLM, LLM and TTS — but it's non-reasoning.
- Ollama is the only documented engine as of release day; we've found no independent benchmarks, so treat the quality claims as the vendor's pitch.

Liquid AI's LFM3 family has landed on Ollama today (29 October 2026), and it's a big drop for small-hardware people: four models — [lfm3](https://ollama.com/library/lfm3) (text), [lfm3-vl](https://ollama.com/library/lfm3-vl) (vision), [lfm3-audio](https://ollama.com/library/lfm3-audio) and [lfm3-omni](https://ollama.com/library/lfm3-omni) — all released together, all under the LFM Open License v1.0, and all sized for phones, laptops and edge devices rather than gaming rigs.

The readmes mention the wider family also includes base and decision models, but those four are the ones you can actually pull today. And there's an architectural story here too: this isn't a plain transformer.

## What the LFM3 family is

Liquid AI has been building hybrid architectures for a while — mixing convolutions and attention rather than stacking attention layers end to end. LFM3 pushes that further into "tri-hybrid" territory. The 1.7B text model's 24 layers break down as:

- 13 gated short convolutions
- 5 sliding-window attention layers
- 6 global attention layers

The pitch is that convolutions are cheap for local patterns, sliding-window attention handles medium-range structure, and only six layers pay the full cost of global attention. That's how you get a 256K-token context window in a 1.73B model that takes about 1 GB of memory at 4-bit, per the readme.

The whole family is edge-first — Liquid AI says every checkpoint is designed for phones, laptops and small devices. Where recent releases like [Bonsai 2](/blog/bonsai-2-on-ollama-a-27b-model-that-fits-in-8-3gb) squeeze a big model into a desktop, LFM3 goes the other way: small models designed for phones from the start.

## The four models at a glance

| Model | Parameters | Download size | Context (readme) | Inputs | What it's for |
|---|---|---|---|---|---|
| lfm3 | 1.73B | 1.1–1.7GB | 256K | Text | Local agents, tool calling |
| lfm3-vl | 1.92B | 1.5–1.7GB | 128K | Text, image | On-device visual understanding |
| lfm3-audio | 2B | 1.7–2.1GB | 32K | Text or audio | Voice chat, transcription, translation |
| lfm3-omni | 2.17B | 2.0–2.4GB | 32K | Text, image, video, audio | One model replacing ASR + VLM + LLM + TTS |

One discrepancy worth flagging: Ollama's variant tables list 125K context for the vl, audio and omni tags, while the readmes say 32K (128K for vl). We'd trust the readme figures until the pages settle — check the live page before relying on the larger number.

## lfm3: the 1.7B text model for agents

`ollama run lfm3`

This is the headline model. It's post-trained for agentic harnesses — the readme names Hermes Agent, Pi and OpenClaw — with tools, thinking and structured outputs listed as capabilities on the Ollama page.

The standout spec is the 256K-token context window (input plus output). That's remarkable at 1.7B, and it changes what a small local agent can do: long documents or lengthy conversation histories fit without the truncation gymnastics you'd normally need at this size. Just remember the memory rule of thumb we use across this blog — budget roughly the download size (1.1–1.7GB) in RAM or VRAM, plus headroom, and the headroom grows with the context you actually use. The full 256K window will need meaningfully more than the weights alone, so keep context modest on small machines.

Other things worth knowing:

- **Made to be specialised.** Liquid AI says quick single-GPU fine-tuning adapts it to your domains, formats and tools. That's an unusual thing to put front and centre — this is a model you're expected to tune, not just run.
- **16 text languages**, including English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Hindi, Russian and Polish.
- **Honest limits, straight from the readme:** not recommended for agentic coding or complex multi-step reasoning, and you should use retrieval or search for knowledge-heavy tasks. We appreciate a vendor that says this on day one.

(Live on Inpherio: https://inpherio.co.uk/models/lfm3:latest)

## lfm3-vl: vision on a 2B budget

`ollama run lfm3-vl`

lfm3-vl pairs the LFM3 language model (1.73B) with a new in-house vision encoder (164M) plus a 25M projector — 1.92B total. The encoder is 14 layers: 8 gated depthwise 2D convolutions and 6 bidirectional attention layers, working on 16×16 patches with 2D rotary embeddings. Images are processed at native resolution, with each 2×2 block of patches becoming one language token.

Two features stand out:

- **Thinking.** The model reasons step by step before answering, which the readme says helps most on math, charts and multi-step visual questions. Thinking is listed as a capability on the Ollama page.
- **Spatial and embodied reasoning.** It finds objects and reasons about how they're placed relative to each other — pitched explicitly as a foundation for robotics applications.

Video input is listed as "soon" — this will be the first LFM vision model to take video, sampled at 2 fps, but it's not there on day one.

The readme is again blunt about limits: not recommended for coding, multi-turn conversations or complex reasoning. This is a document/chart/screen/object model, not a general chatbot. Context is 128K per the readme.

(Live on Inpherio: https://inpherio.co.uk/models/lfm3-vl:latest)

## lfm3-audio: voice chat, transcription and translation in 2B

`ollama run lfm3-audio`

lfm3-audio takes text or 16 kHz audio in, and answers in text, 24 kHz speech, or both — with causal audio encoding, incremental prefill and streaming speech generation. In practice that means you can ask a question aloud and get a streamed reply in text and speech.

What it covers:

- **Voice chat**, with reasoning before answering for text-only responses.
- **Multilingual speech:** transcribe, synthesise and translate across 18 languages and locales — including British and American English, Danish, Greek, Finnish, Hebrew, Swedish and Turkish.
- **Audio understanding and reasoning:** interpret recordings and answer questions about the speech, sounds and events in them.
- **Tool calling from spoken requests** — "turn on the lights" becomes a tool call, with text or speech alongside it.

It introduces a new in-house audio encoder and the LiquidCodec audio tokenizer, which Liquid AI says improves speech quality over LFM2.5-Audio. Context is 32K per the readme — fine for voice chat, tight for long transcripts.

The same limits apply as the rest of the family: not for coding or complex multi-step reasoning, and use retrieval for knowledge-heavy tasks.

## lfm3-omni: one model instead of four

`ollama run lfm3-omni`

This is the most novel of the family, and the one to watch. lfm3-omni is Liquid AI's first omni-modal model: a single end-to-end model that takes text, images, video (with or without its soundtrack) and 16 kHz audio, and answers in text, 24 kHz speech, or both.

The pitch is consolidation. Instead of running a separate ASR model, vision-language model, LLM and TTS model on a device, you run one 2.17B model. On memory-constrained edge hardware, where every resident model costs you something, that's a genuinely different proposition from bolting four models together.

The interesting capability is that it **watches and listens together**: for video with sound, it processes both simultaneously, so it can answer questions that need sight and hearing at once — "what made that noise?" about a clip, say. Images and video are processed at native resolution (each 2×2 patch block becomes one language token), video sampled at 2 fps.

One caveat worth flagging clearly: this is a **non-reasoning model**. Thinking is not supported — unlike lfm3 and lfm3-vl. If you want step-by-step reasoning, the omni model isn't it.

It's a 2.0–2.4GB download, designed for phones, laptops and edge hardware, with 32K context per the readme.

## How to run them and what hardware you need

All four are one-liners:

```bash
ollama run lfm3
ollama run lfm3-vl
ollama run lfm3-audio
ollama run lfm3-omni
```

Each has a single size tag (1.7b for lfm3, 2b for the rest), so `latest` and the size tag are the same download — no quant-picking required.

**Memory budgeting.** As we covered in [LM Studio vs Ollama](/blog/lm-studio-vs-ollama-which-one-should-you-actually-install), the rule of thumb is to budget roughly the download size in RAM/VRAM plus context headroom. So lfm3 wants around 1.1–1.7GB plus context; lfm3-omni wants 2.0–2.4GB plus context. The bigger the context window you actually fill, the more extra memory you need on top.

**Hardware fit.** Every model here is 2.4GB or less, and Liquid AI's design target is phones, laptops and edge devices. In practice that means a laptop with 8GB RAM should handle all four at modest context, and any Mac or gaming GPU has plenty of room to spare. This is the rare family where the hardware question is "do you have a computer", not "which GPU" — just don't expect to fill a 256K context window on the smallest machines.

**Engines.** As of release day, the Ollama library pages are the only documentation we've seen — we found no LM Studio, llama.cpp or vLLM instructions on them. That's the same situation we found with [kev](/blog/kev-on-ollama-jared-palmer-s-decision-model-family-hands-on) and other day-one releases, and it usually resolves within weeks as GGUF files spread. For now, Ollama is the way to run these.

If you want to serve them rather than just run them, the [connector](/connector) supports Ollama — and [new hosts pay no platform fee for their first 12 months, up to £250 of earnings](/providers).



## What it's good for — and what to reach for instead

**Good for:**

- Local agents and tool calling on small devices (lfm3)
- On-device document, chart and screen reading, and object finding (lfm3-vl)
- Voice interfaces, transcription and translation (lfm3-audio)
- Any project where one small model must do several jobs at once (lfm3-omni)

**Not for:** agentic coding, complex multi-step reasoning, and knowledge-heavy tasks without retrieval — the readmes say so explicitly for all four. Don't buy the hype in either direction; a 2B model is a 2B model.

**How it compares.** In the ~2B text class, the familiar rivals are [qwen3:1.7b](/models/qwen3:1.7b) and the small Gemma variants. lfm3's differentiators are the 256K context, the tri-hybrid architecture and the agentic post-training. For small local vision, [moondream:1.8b](/models/moondream:1.8b) is an established option; lfm3-vl adds thinking and spatial reasoning, but it's a day old with no third-party numbers.

And that's the honest caveat for the whole family: everything on the pages is Liquid AI's own pitch, and we've found no independent benchmarks yet. If your tasks exceed what a 2B model can do, an 8B-class model like [qwen3:8b](/models/qwen3:8b) is the next step up — with a correspondingly larger memory bill.

**What we'd pick:** try `ollama run lfm3` first — the 256K context at 1.7B is the most immediately useful trick in the family — and lfm3-omni if you're building anything with voice and images on constrained hardware. Give the others a week or two for the community numbers to arrive before committing to them in anything serious.

## FAQ

**What is lfm3 on Ollama?** lfm3 is Liquid AI's 1.73B text model for local agents and tool calling, released on Ollama on 29 October 2026 as part of a four-model LFM3 family, under the LFM Open License v1.0.
**How much memory does lfm3 need?** The download is 1.1–1.7GB, and Liquid AI says about 1 GB at 4-bit precision. Budget roughly the download size in RAM/VRAM plus headroom for the context you actually use.
**Can lfm3 run on a laptop or phone?** Yes — that's the design target. Every LFM3 model is 2.4GB or under, so a laptop with 8GB RAM should handle all four at modest context, and any Mac or gaming GPU has room to spare.
**What's the difference between lfm3, lfm3-vl, lfm3-audio and lfm3-omni?** lfm3 is text-only for agents; lfm3-vl adds image understanding; lfm3-audio does speech in and speech out; lfm3-omni takes text, images, video and audio and answers in text or speech.
**Does lfm3 support tool calling?** Yes — tools and thinking are listed capabilities, and it's post-trained for agent harnesses like Hermes Agent, Pi and OpenClaw.
**Is lfm3 good for coding?** No. The readme says it's not recommended for agentic coding or complex multi-step reasoning due to its small size; use retrieval for knowledge-heavy tasks.
