LFM3: Liquid AI's Model Family, Sizes, Benchmarks, How to Run

LFM3 is Liquid AI's new small-model family, built to run on your own device and to be fine-tuned for one job. Liquid AI announced the family on Oct 27, 2026 with a new "tri-hybrid" architecture and new multimodal encoders. Every model is open-weight and live on Hugging Face.

Here is what LFM3 is, the six models, the benchmarks, the license, and how to run it today.

1. What is LFM3?

The name stands for Liquid Foundation Models, generation 3. It follows LFM2 (July 2025) and LFM2.5. Liquid AI, an MIT spin-off in Cambridge, Mass., designs these models for phones, laptops, cars and edge chips instead of the cloud.

The big change is one shared backbone. Liquid trained the 1.7B base on 22T tokens, then branched text, vision, audio, omni and decision models from it. You pick the model that fits your task, fine-tune it on one GPU, and ship it on-device.

2. LFM3 models and sizes

There is also LFM3-1.7B-Base for your own fine-tunes, plus official GGUF, MLX, ONNX, FP8 and NVFP4 builds. The GGUF build of the text model already passed 170K downloads on Hugging Face.

3. LFM3 architecture: the tri-hybrid design

• 24 layers: 13 gated short convolutions, 5 sliding-window attention layers (256-token window), 6 global attention layers.
• Head-wise gated attention and partial RoPE to keep long context cheap and quantization clean.
• 128K vocabulary shared with LFM2.5-8B-A1B, better for non-Latin scripts.
• Speed: at 8K tokens of context, LFM3-1.7B is about 25% faster per parameter than LFM2-1.2B. A full 128K context needs only 1.6 GB of KV cache.

4. LFM3 benchmarks

Liquid AI tested the text model against models up to 9.7B parameters. The text model wins on agents and instruction following, and loses on knowledge and coding, where size matters most.

LFM3-VL-2B averages 71.4 across 62 vision tests, close to Gemma 4 12B (71.7), a model six times bigger. LFM3-Omni-2B scores 70.1 on Daily-Omni versus 61.3 for Qwen2.5-Omni-7B. All scores are from Liquid AI's launch post.

5. LFM3 license

The family ships under the LFM Open License v1.0. You can download, modify and fine-tune the models for free. Commercial use is licensed for companies under $10 million in annual revenue; bigger companies need a separate deal with Liquid AI. Non-profits doing research are exempt from the cap.

6. How to run LFM3

• Ollama: ollama run lfm3 (also lfm3-vl, lfm3-audio, lfm3-omni).
• llama.cpp / LM Studio: use the official LFM3-1.7B-GGUF; the smallest build needs about 1 GB of RAM.
• Apple Silicon: MLX builds in 4, 6 and 8 bit.
• Servers: vLLM and SGLang; on one NVIDIA B300, LFM3-1.7B produces about 3.6B tokens a day.
• Cloud: Liquid AI API, Playground, and OpenRouter (a launch partner).
• Settings: temperature 0.1, top_k 50, repetition_penalty 1.1 for chat and agents. Reasoning is on by default.

The text model is also post-trained for agent harnesses like Hermes Agent, Pi and OpenClaw, so a local agent can triage email or run tool calls with no cloud bill.

7. LFM3 vs LFM2

LFM2 searched architectures on one phone and one laptop CPU; LFM3 searched across 14 device and backend pairs, from a Raspberry Pi 5 to Qualcomm NPUs. The new design doubles the convolution kernel (7 vs 3), adds sliding-window attention, and stretches context from 32K to 256K. For agent work, LFM3-1.7B beats the bigger LFM2.5-2.6B on almost every test.

8. Where Sumus fits

Yes, you can put a small model like LFM3 to work and still get frontier answers when you need them. A Sumus AI Employee sits in your team chat, takes a task in plain words, and picks the right model for each step, from fast small models to Claude and GPT. Ask it to compare LFM3 with Qwen3.5 on your own prompts and post the results back to the thread.

Sources

• Liquid AI: Introducing LFM3, Oct 27, 2026
• LFM3 on Hugging Face
• Liquid AI on X, Oct 27, 2026; LFM3-1.7B model card and LICENSE, read Nov 3, 2026