What Is Liquid AI LFM3? Six Open-Weight 2B Models Explained
The short answer
LFM3 is a family of six small open-weight models from Liquid AI, released October 27, 2026, for running specialised AI on-device. Instead of one general model, you pick the model for the input you have — text, images and video, audio, or all of them — fine-tune it on one GPU and run it locally. Weights are on Hugging Face, with GGUF, MLX, ONNX, FP8 and NVFP4 builds, and the models are also in the Liquid AI API and Playground.
| Model | Parameters | Inputs | Use |
|---|---|---|---|
| LFM3-1.7B | 1.73B | Text | Local agents, tool calling, structured output |
| LFM3-VL-2B | 1.9B | Text, images, video | Vision QA, documents, GUI understanding |
| LFM3-Audio-2B | ~2B | Text, audio | Voice chat, speech-to-text, text-to-speech in 18 languages |
| LFM3-Omni-2B | 2.2B | Text, images, audio, video | Spoken visual QA, multimodal tool calls |
| d1-2B | ~2B | Text, images | Fast typed decisions (classify, point, extract) |
| d1-Omni-2B | ~2B | Text, images, audio | Decisions from voice input |
What is new in the architecture
LFM3 keeps LFM2’s idea — mostly short convolutions instead of attention — and adds attention where long context needs it. Of 24 layers, 13 are gated short convolutions, 5 are sliding-window attention with a 256-token window and 6 are global attention. The base model was pre-trained on 22 trillion tokens, the last 3.5 trillion a mid-training stage with more agentic traces and tool calls that also extended context to 256K tokens. Liquid AI says that at 8K tokens of context LFM3-1.7B runs about 25% faster than LFM2-1.2B on edge devices.
How good is it?
On Liquid AI’s own benchmark table (vendor figures, October 27, 2026), LFM3-1.7B:
- beats much larger models on agent tasks: LiveClaw 45.6 against 35.8 for Qwen3.5-9B, WebArena-Verified 38.4 against 28.9, AutomationBench 41.3 against 15.1;
- leads on function calling: BFCL V4 71.4, ToolSandbox 80.4;
- loses on knowledge and coding: GPQA Diamond 63.4 against 80.4 and SWE-bench Verified 43.4 against 51.2 for Qwen3.5-9B.
LFM3-VL-2B averages 71.4 across 62 vision benchmarks, which Liquid AI puts level with Gemma 4 12B (71.7). MacPaw reported that LFM3-1.7B completed 64% of 65 everyday macOS agent tasks against 48% for Qwen3.5-2B, using 58% fewer tokens. Independent evaluations had not been published on October 28, 2026.
Licence: free below $10 million revenue
The models use the LFM Open License v1.0, an Apache-style licence with one important condition: commercial use is licensed only if your organisation’s annual revenue is under $10 million. Above that, you need a commercial agreement with Liquid AI. Check this before building a product on it; it is not an open-source licence in the OSI sense.
How to run it
The model cards list vLLM, SGLang and Transformers for GPUs, and llama.cpp, Ollama and LM Studio for local use; LFM3-1.7B is post-trained to work inside agent harnesses such as Hermes Agent, Pi and OpenClaw. Use an LFM3-compatible llama.cpp build with the current chat template. To embed it in an app, see best on-device LLM runtimes.
Related: for earlier decision models, see OpenAI Decisions API vs Microsoft Decision-1 vs Liquid d1; for hosted model prices see current API prices.
Last verified: October 28, 2026.