Liquid AI Releases LFM3, Six Open-Weight Models Around 2B Parameters Covering Text, Vision, Audio and Omni Inputs
Liquid AI said on October 27 that the first release of its LFM3 family covers six models in the 2B-parameter range, all trained from LFM3-1.7B-Base: a text model, a vision-language model, an audio model, an omni model and two decision models. Base and post-trained weights are on Hugging Face under Liquid's own LFM Open License v1.0, which limits commercial use to organizations with less than $10 million in annual revenue. The benchmark comparisons are Liquid's own.
Liquid AI's LFM3 release is aimed at phones and laptops rather than data centers, and the claim worth testing is that a model of under 2 billion parameters can run agent tasks well enough to be useful on-device. The weights are open, but the license is not a standard open-source one: larger companies need a separate arrangement for commercial use.
What Liquid released#
Liquid AI said on October 27 that "the first release of the LFM3 family covers models in the 2B parameter range." It lists six models, all trained from LFM3-1.7B-Base:
- LFM3-1.7B, a text model with a 256,000-token context window, post-trained for agent harnesses.
- LFM3-VL-2B, a 1.9-billion-parameter vision-language model that also accepts video input.
- LFM3-Audio-2B, an audio model covering 18 languages.
- LFM3-Omni-2B, a 2.2-billion-parameter model that Liquid calls its first omni model.
- d1-2B and d1-Omni-2B, which Liquid describes as decision models.
Liquid says base and post-trained weights are available on Hugging Face, and that the models can also be used in its Playground and through the Liquid AI API.
Liquid says the models were pretrained on 22 trillion tokens using what it calls a tri-hybrid architecture. It says LFM3-1.7B takes about 1 GB of memory at 4-bit precision and that every model can be fine-tuned on a single GPU.
Liquid's benchmark claims#
Liquid says LFM3-1.7B outperforms its previous LFM2.5-2.6B with 35% fewer parameters, and reports an average score of 49.37 across its chosen benchmarks against 47.24 for Qwen3.5-9B. It reports 38.42 on WebArena-Verified, a web-agent benchmark, against 3.02 for LFM2.5-2.6B. Liquid also says larger models still do better on agentic coding and knowledge-heavy tasks.
For vision, Liquid says LFM3-VL-2B averages 71.4 across 62 benchmarks, against 71.7 for Gemma 4 12B, and is within two points of GPT-5.6 Luna at 73.4.
Liquid names Ambarella, AMD, Arm, Intel, NVIDIA, OpenRouter, Qualcomm and Vercel as launch partners, and says Mercedes-Benz and Shopify tested the models before release.
The license#
The LFM3-1.7B model card lists the LFM Open License v1.0. The license grants broad rights to use, modify and redistribute the weights, but says commercial use is licensed only for organizations below a threshold of $10 million in annual revenue. The threshold does not apply to qualifying nonprofits using the models for non-commercial or research purposes.
Limits#
- All benchmark numbers are Liquid's own, on benchmark sets it selected. We found no independent evaluation yet.
- We checked the license on the LFM3-1.7B model card only; we did not open the license file for each of the six models.
- Liquid uses "open-weight" for these models. Under its license terms they are not open source in the usual sense for larger companies.