What happened
Liquid AI announced LFM3 on October 27, 2026, a new generation of its Liquid Foundation Models. All six first models branch from one LFM3-1.7B-Base backbone, pre-trained on 22T tokens with a new tri-hybrid architecture, and sit around 2B parameters.
The base and post-trained models are open-weight under the LFM Open License v1.0 and available on Hugging Face, Liquid's Playground and the Liquid AI API. On Hugging Face, Liquid AI lists GGUF, ONNX and MLX builds for all six models, d1-2B and d1-Omni-2B included.
The six models
Liquid AI lists four modality models and two decision models:
- LFM3-1.7B: text, 256K-token context, about 1 GB at 4-bit precision.
- LFM3-VL-2B: vision, now with video input for the first time in an LFM model.
- LFM3-Audio-2B: speech with in-house encoder and codec, 18 languages.
- LFM3-Omni-2B: video, audio and text together, including multimodal tool calling.
- d1-2B (built on LFM3-VL-2B) and d1-Omni-2B (built on LFM3-Omni-2B): decision models that return typed answers in milliseconds instead of text.
What it means for you
Liquid AI claims LFM3-1.7B outperforms its previous LFM2.5-2.6B with 35% fewer parameters, and that LFM3-VL-2B leads similar-sized models by at least 7.4 points across 62 vision benchmarks. These are the company's own results.
Every model can be fine-tuned on a single GPU, according to Liquid, and runs with llama.cpp, MLX and vLLM. Liquid's docs library still marks GGUF for the audio and d1 models as coming soon, although the files are already on Hugging Face. Hosted API prices are not established on the public announcement. Day-zero launch partners named by Liquid include AMD, Arm, NVIDIA, Qualcomm, Vercel and OpenRouter.
Go to the source
Sources & further reading
Based on the official documentation below. Features and access can change; check the provider before making a decision.
Was this page useful?
Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.
Go a little further