Liquid AI LFM3: How to Run the 2B Text, Vision and Audio Models Locally (Ollama, llama.cpp, LM Studio)

Most “small” local models still want 8 GB of VRAM or more. LFM3, released by Liquid AI on October 27, 2026, goes the other way: six open-weight models of about 2B parameters each, built to run on laptops, phones and mini PCs. The text model, LFM3-1.7B, needs about 1 GB at 4-bit, handles a 256K-token context, and beats Qwen3.5 9B on several agent and tool-calling benchmarks. There are also vision, audio and omni versions. This guide explains the six models, how they compare, and how to run them locally with Ollama, llama.cpp and LM Studio, with the exact commands and settings.

What Is LFM3?

LFM3 is the third generation of Liquid AI’s “Liquid Foundation Models”. Instead of one big model, it is a family that shares a single 1.7B-parameter backbone and adds small encoders for vision and audio. Liquid AI aims it at on-device and specialised AI: local agents, document reading, voice assistants and quick classification jobs that should not need a cloud API.

The weights are on Hugging Face in several formats: GGUF for llama.cpp, Ollama and LM Studio, MLX for Apple Silicon, ONNX, and FP8 and NVFP4 for server GPUs. A base model, LFM3-1.7B-Base, is also available for your own fine-tuning.

The Six LFM3 Models

For most readers, the two to try first are LFM3-1.7B as a local assistant or agent brain, and LFM3-VL-2B for reading screenshots and documents. The d1 models are more specialised. They return calibrated, typed answers for automation pipelines rather than chatting.

Why LFM3 Is So Small and Fast

LFM3 is not a standard transformer. Its 24 layers are mostly gated short convolutions, which are cheap to run on CPUs and phones, with attention added only where long context needs it:

• 13 gated short convolution layers (kernel size raised from 3 to 7)
• 5 sliding-window attention layers with a 256-token window
• 6 global attention layers, using partial RoPE

Because only 6 layers keep a full attention cache, the KV cache stays small even at long context. That is why a 1.7B model can offer 256K tokens without eating all your RAM. Liquid AI chose this layout with hardware-in-the-loop search across 14 device and backend combinations. The base model was pre-trained on 22 trillion tokens, with a final 3.5 trillion-token stage full of agent traces and tool calls, which explains its agent scores.

Liquid AI says LFM3-1.7B prefills at least 2.5x faster and decodes at least 1.8x faster than the compared models at 32K context on its test devices. These are vendor claims.

LFM3-1.7B Benchmarks

Here is how the 1.7B text model compares, using Liquid AI’s own published scores:

The pattern is clear. For its size, LFM3-1.7B is excellent at following instructions, calling tools and acting as an agent, beating a 9B model in those areas. It is weaker than larger models at deep reasoning, agentic coding and very long-context recall. Liquid AI itself says it is not recommended for agentic coding or complex multi-step reasoning, and suggests pairing it with search or retrieval for knowledge-heavy questions. One third-party data point: MacPaw reported LFM3-1.7B completed 64% of 65 everyday macOS agent tasks, against 48% for Qwen3.5-2B, while using 58% fewer tokens.

No independent benchmark results had been published at the time of writing.

Estimated speed: a 1.7B model at Q4_K_M is small enough that a modern 16 GB GPU such as an RTX 5080 should generate well over 100 tokens per second, and a recent 8-core desktop CPU roughly 20 to 50 tokens per second. These are estimates based on the model size, not measured results.

Hardware and File Sizes

At about 1.1 GB for Q4_K_M, LFM3-1.7B fits on any GPU from the last decade, on Apple Silicon Macs, and runs well on a CPU alone. Q8_0 at 1.85 GB is a good choice if you have the room, since quality loss at 8-bit is minimal. The full F16 file is only worth it for fine-tuning or exact reference testing.

Compared with other small local models we have covered, such as Nemotron 3.5 Bolt E4B (about 4 GB) and MiniCPM-V 4.7, LFM3 is the smallest by a wide margin. That makes it a strong pick for always-on background agents that should not take memory away from ComfyUI or games.

Run LFM3 With Ollama

You need Ollama 0.40.12 or newer, because older versions do not know the lfm3 architecture. Check with ollama --version and update from ollama.com if needed.

ollama run hf.co/LiquidAI/LFM3-1.7B-GGUF:Q4_K_M

This pulls the GGUF directly from Hugging Face. Swap the tag for :Q8_0 or :Q5_K_M to get another quantization. The model card also lists ollama run lfm3 as an option from the Ollama library.

Ollama stores models in its own folder: ~/.ollama/models on macOS and Linux, and C:\Users\YourName\.ollama\models on Windows. You do not need to move any files yourself.

To use it from other apps, Ollama serves an OpenAI-compatible API at http://localhost:11434/v1. Set the temperature to 0.1 in your client, because the default is much higher (see the settings section below).

Run LFM3 With llama.cpp

You need llama.cpp build b11960 (v0.8.0) or newer. Download a release from the llama.cpp releases page or build from source. These are the commands from the model card:

# Chat, tools and agents
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M --temp 0.1 --top-k 50 --repeat-penalty 1.1

# Math and science
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M --temp 1.0 --top-p 0.95 --top-k 20

# The separate QAD Q4_0 build
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF --hf-file LFM3-1.7B-QAD-Q4_0.gguf --temp 0.1 --top-k 50 --repeat-penalty 1.1

# Local OpenAI-compatible server
llama serve -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M

The -hf flag downloads the file to llama.cpp’s cache automatically. If you already downloaded a GGUF by hand, use -m path/to/your-model.gguf instead. The card also notes you can load a custom chat template with --chat-template-file if your build’s built-in template is out of date, a common cause of strange output with brand-new architectures.

Run LFM3 in LM Studio

1. Update LM Studio and make sure its llama.cpp engine is version 2.65.1 or newer (Settings, Runtime).
2. Search for LiquidAI/LFM3-1.7B-GGUF in the Discover tab, or open the model page on Hugging Face and choose Use this model, LM Studio.
3. Pick Q4_K_M or Q8_0 and download.
4. In the chat settings, set temperature 0.1, top K 50 and repeat penalty 1.1.

LM Studio also has a local server mode that exposes an OpenAI-compatible API, so you can plug LFM3 into tools like Open WebUI or your own scripts.

Run LFM3-VL-2B for Images and Video

LFM3-VL-2B adds a 164M-parameter vision encoder to the 1.7B backbone. It reads images at native resolution and samples video at 2 fps (4 to 768 frames). Its context is 128K tokens. On Liquid AI’s 62-benchmark average it scores 71.4, close to Gemma 4 12B’s 71.7, and it is particularly strong on documents and screens: 96.7 on DocVQA, 92.7 on ChartQA and 94.6 on ScreenSpot-v2 Desktop. It is weaker on MMMU (55.7) and Video-MME (62.8) than larger models.

With llama.cpp b11960 or newer, start a server (the vision projector, an F16 mmproj file, downloads automatically with -hf):

llama serve -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_0 --temp 0.1 --top-p 1.0 --top-k 50 --min-p 0.0 --repeat-penalty 1.1 --no-reasoning-preserve --port 8080

For a single question, the card uses llama cli with the same flags plus --image path/to/image.png or --video path/to/clip.mp4 and -p "your question". Video input needs ffmpeg on your PATH, and in llama.cpp the 2 fps sampling is fixed (--video-fps has no effect).

Good uses: reading receipts and forms, describing screenshots for automation, checking charts, and captioning images for training datasets in ComfyUI workflows.

LFM3-Audio and LFM3-Omni: Local Voice

LFM3-Audio-2B is a complete local voice stack in one model: speech recognition (6.23% English WER), text-to-speech at 24 kHz in 18 languages with a female or male voice, English voice chat, speech translation and spoken tool calls. Liquid AI quotes a time to first token as low as 50 ms, and its LiquidCodec decodes speech at about 87x real time on a single CPU thread of an M4 Max. A live, chunked English transcription mode is included.

LFM3-Omni-2B combines everything: it takes text, images, video with sound, and audio, and answers in text, speech or both. Its context is 32K tokens, and it does not support thinking mode.

Both have GGUF builds (LiquidAI/LFM3-Audio-2B-GGUF and LiquidAI/LFM3-Omni-2B-GGUF) for llama.cpp, including a realtime speech server, plus MLX and ONNX versions of the Audio model. Liquid AI’s llama.cpp deployment docs have the current commands. For Python, install:

pip install "transformers>=5.20.2" torch accelerate librosa soundfile

Then load the model with AutoModelForMultimodalLM.from_pretrained("LiquidAI/LFM3-Audio-2B", trust_remote_code=True) and its processor. For text-to-speech, the system prompt format is “Perform TTS in English using a female voice.” If you want more voice tools, see our Voicebox local voice guide.

Best Settings and Thinking Mode

Thinking is on by default for LFM3-1.7B and VL-2B. The chat template ends its generation prompt with <think>, so the model reasons first. You can turn it off with enable_thinking=False in Transformers, but Liquid AI warns this can be inconsistent on reasoning-heavy tasks. Reasoning from earlier turns is stripped from the history unless you set preserve_thinking=True. Allow plenty of output tokens; the VL example uses up to 16,384.

The most common mistake is running LFM3 at a default temperature of 0.7 or 0.8. Small models drift and repeat at high temperatures, so stick to 0.1 for anything except maths.

Tool Calling and Agents

Tool calling is LFM3-1.7B’s best skill. Pass your function schemas through the chat template’s tools argument. The model replies with XML-style calls: a <tool_call> block containing <function=NAME> and <parameter=NAME> tags. Parse it, run the tool, add the call to history as a tool_calls entry and the result as a tool message. One gotcha from the card: function.arguments must be a Python dict, so run JSON strings through json.loads() first.

It also supports deferred tools: set defer_loading=True and tool definitions stay out of the prompt until a search tool returns their names, which saves context when you have many tools. Liquid AI post-trained the model to work inside the Hermes Agent, Pi and OpenClaw harnesses (agent harness docs).

A practical setup: run LFM3-1.7B in Ollama as an always-on helper that sorts files, fills forms or drives simple automations, and hand anything that needs deep reasoning to a bigger model. For a look at what the giant end of open models needs, see our Mistral Large 4 open weights guide.

Licence: Can You Use It Commercially?

LFM3 uses the LFM Open License v1.0. It is Apache-style for most uses, but it is not an OSI open-source licence. The key condition: commercial use is allowed only if your organisation’s annual revenue is under $10 million. Above that, you need a commercial agreement with Liquid AI. Personal use, research and small businesses are covered. Read the full licence text before shipping a product.

FAQ

How much RAM does LFM3 need?

The LFM3-1.7B Q4_K_M file is about 1.1 GB, and Liquid AI quotes about 1 GB of memory at 4-bit. Add some headroom for context. Any PC with 4 GB of free RAM, or any GPU, can run it.

Can LFM3 run on a CPU?

Yes. Its convolution-heavy design was built for CPUs, phones and laptops, and llama.cpp and Ollama run it on CPU without a GPU.

Is LFM3 better than Qwen3.5?

For tool calling, agents and instruction following, LFM3-1.7B beats Qwen3.5 4B and even 9B on Liquid AI’s benchmarks. For reasoning, coding and long-context recall, Qwen3.5 9B is stronger.

Why does Ollama say the model architecture is unknown?

Your Ollama is too old. Update to 0.40.12 or newer. For llama.cpp, use build b11960 or newer; for LM Studio, engine 2.65.1 or newer.

Does LFM3 support images?

The text model does not. Use LFM3-VL-2B for images and video, or LFM3-Omni-2B for images, video and audio together.

Can I fine-tune LFM3?

Yes. Liquid AI says it can be fine-tuned on a single GPU, and its cookbook notebooks cover SFT, DPO, GRPO and continued pre-training with TRL. Start from LFM3-1.7B-Base for custom post-training.

Final Thoughts

LFM3 shows how much a well-designed 2B model can do. LFM3-1.7B is a genuinely useful local agent at around 1 GB, LFM3-VL-2B reads documents and screens at near 12B quality, and the Audio and Omni models put a full voice assistant on a laptop. It will not replace a big reasoning model, but as a fast, always-available helper that leaves your GPU free for ComfyUI, it is one of the best options out there right now. Start with the one-line Ollama command, keep the temperature at 0.1, and give it a tool or two to call.

More local AI guides:

• Nemotron 3.5 Bolt E4B: 4 GB GGUF on an RTX 5080
• MiniCPM-V 4.7: Run It Locally
• Mistral Large 4 Open Weights

Sources

• Hugging Face: LiquidAI/LFM3-1.7B and LFM3-1.7B-GGUF (model cards, read October 31, 2026)
• Hugging Face: LiquidAI/LFM3-VL-2B
• Hugging Face: LiquidAI/LFM3-Audio-2B and LFM3-Omni-2B
• Hugging Face: LFM3 collection
• Liquid AI blog: LFM3
• andrew.ooo: What Is Liquid AI LFM3? (licence summary and MacPaw result)