LFM3
Liquid AI's family of 2B-class open models for on-device agents, vision, audio and decisions
Context window
256K
Input / 1M tokens
Free
Output / 1M tokens
Free
Provider
Liquid AI
Open weights on Hugging Face under the LFM Open License v1.0. Liquid AI's launch post lists no hosted API pricing. · Data verified 2026-11-03
LFM3 (October 27, 2026) is Liquid AI's third-generation family of small open-weight models, all built from LFM3-1.7B-Base. It includes LFM3-1.7B for text, agentic use and tool calling with a 256K-token context window, LFM3-VL-2B for image and video input, LFM3-Audio-2B and LFM3-Omni-2B for speech in and out, and d1-2B decision models that return calibrated answers instead of generated text. Liquid AI reports that LFM3-1.7B outperforms the larger LFM2.5-2.6B, and its model card advises against using it for agentic coding or complex multi-step reasoning.
How to access it
Download open weights from Hugging Face (for example LiquidAI/LFM3-1.7B) under the LFM Open License v1.0. Liquid AI points to llama.cpp for on-device vision inference and provides fine-tuning notebooks.
Strengths
- ✓Small enough for local and on-device use (about 1.7B to 2B parameters)
- ✓256K-token context window on the text model
- ✓Text, vision, audio and omni-modal variants from one base model
- ✓38.42 on WebArena-Verified for LFM3-1.7B vs 3.02 for LFM2.5-2.6B (Liquid AI-reported)
- ✓Base model released for custom post-training
Best for developers who...
When to choose it (and when not to)
Reach for LFM3 when...
- →You need a model that runs locally on modest hardware
- →You want a small base model to fine-tune for one task
Look elsewhere if...
- ✕Agentic coding or complex multi-step reasoning (the model card advises against it)
- ✕Knowledge-heavy tasks without retrieval or search
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| WebArena-Verified | 38.42 | LFM3-1.7B, Liquid AI-reported; LFM2.5-2.6B scored 3.02 |
| Vision benchmark average (62 benchmarks) | 71.4 | LFM3-VL-2B, Liquid AI-reported; LFM2.5-VL-3B scored 57.6 |
| UTMOSv2 speech quality | 3.83 | LFM3-Audio-2B, Liquid AI-reported; LFM2.5-Audio-1.5B scored 3.22 |
Compare LFM3
Compare LFM3 with any other model
Build a comparison →All model comparisons →