Liquid AI announces LFM3: six small models, text to voice
Liquid AI released the LFM3 family on 27 October 2026: text, vision, audio and omni models of about 2B parameters built for devices, plus LFM3-Audio-2B for voice, under the LFM Open License.
Liquid AI announced LFM3 on 27 October 2026 as the next generation of its Liquid Foundation Models. The first release has six models of about 2 billion parameters each, all branched from a single 1.7-billion-parameter base model that was trained on 22 trillion tokens. Base and post-trained weights are open on Hugging Face.
What is in the family
- Text, vision, audio and omni models, plus two “decision” models specialized from the vision and omni ones.
- LFM3-1.7B is the text model. It was post-trained for agent harnesses such as Hermes Agent and OpenClaw, and it has a 256K-token context and supports 16 languages. At 4-bit precision it takes about 1 GB.
- LFM3-VL-2B adds image input and, for the first time in an LFM, video input, together with step-by-step reasoning.
- LFM3-Omni-2B, with 2.17 billion parameters, takes text, images, video and audio and answers in text or speech. Its context is 32K tokens.
- LFM3-Audio-2B is the voice model. It takes text or 16 kHz audio and answers with text, 24 kHz speech or both. It covers voice chat with streaming replies, transcription, speech synthesis and speech translation across 18 languages, audio understanding and turning spoken requests into tool calls. Voice chat currently supports English only. The BF16 checkpoint is about 4 GB, and GGUF, MLX 4-bit and ONNX builds are published too.
The architecture
LFM3 uses a “tri-hybrid” design with three types of layer. LFM3-1.7B has 24 layers: 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. Liquid says it beats its previous LFM2.5-2.6B while using 35% fewer parameters. Audio models use a new in-house causal encoder and an audio tokenizer called LiquidCodec, so the model can start processing while you are still speaking. Liquid reports a time to first token as low as 50 milliseconds.
Read the caveats
Liquid’s own tables place LFM3-1.7B ahead of several larger models on tool use and instruction following, and its speech figures, including an English word error rate of 6.2%, are its own measurements. The model card itself is more cautious: it recommends retrieval or search for knowledge-heavy tasks, and says the model is not recommended for agentic coding or complex multi-step reasoning because of its small size.
License
The LFM Open License v1.0 is Apache-style, with one catch: commercial use is licensed only while your company, together with its affiliates, earns under US$10 million a year. Above that line you need your own agreement. For on-device models from other developers, see small models for laptops.
Sources
- Liquid AI: Introducing LFM3, Every Modality for Specialized Intelligence (27 Oct 2026) liquid.ai
- LFM3-1.7B model card (Hugging Face) huggingface.co
- LFM3-Omni-2B model card (Hugging Face) huggingface.co
- LFM3-Audio-2B model card (Hugging Face) huggingface.co
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| LFM3 Liquid AI | 1.73B – 2.17B | 256K | LFM Open 1.0 (Conditional) | 1.19 GB Laptop or phone | |
| LFM3-Audio-2B Liquid AI | 2B | 32K | LFM Open 1.0 (Conditional) | 4.6 GB Laptop or phone |