Liquid AI Releases LFM3: Six Open-Weight 2B-Class Multimodal Models Built on a Single Backbone
Liquid AI has released LFM3, its new generation of Liquid Foundation Models. LFM3 is a family of six open-weight, 2B-class models built on one shared backbone, covering text, vision, audio, omni, and decision tasks. Every model can be fine-tuned on a single GPU and then deployed on phones, laptops, private servers, or the cloud—a positioning that fits squarely into the 2026 shift toward capable, deployment-ready small models at the edge.
TL;DR
- Size: LFM3-1.7B has 1.73B parameters and a 256K context window. LFM3-VL-2B has 1.92B parameters and a 128K context. LFM3-Omni-2B has 2.2B parameters.
- Runs on: Roughly 1 GB of memory at 4-bit quantization, making it viable for phones and laptops. GGUF, MLX, ONNX, FP8, and NVFP4 checkpoints are available.
- Performance: LFM3-1.7B averages 49.37 across 29 benchmarks, ahead of Qwen3.5-9B at 47.24 (per Liquid AI's own evaluation).
- Strongest results: IFEval at 96.66, and AutomationBench at 41.29 versus 15.10 for Qwen3.5-9B.
- Weakest results: Humanity's Last Exam (no tools) at 7.17, and RULER 128K at 65.56 versus 87.11 for Qwen3.5-9B.
What Is Liquid AI LFM3?
LFM3 is a multimodal model family from Liquid AI that branches six models from a single base. All of them start from LFM3-1.7B-Base, which keeps the training and fine-tuning pipeline consistent across modalities. The shared backbone approach lets developers adopt one architecture, toolchain, and quantization format while selecting the variant that matches their task—text, vision, audio, omni, or decision-making workloads.
With sub-2B parameter variants, aggressive quantization support, and broad runtime compatibility (GGUF, MLX, ONNX, FP8, NVFP4), LFM3 targets the growing demand for models that run locally on commodity hardware without a GPU cluster. That makes the family well suited to on-device assistants, private enterprise deployments, and agentic workflows where latency and data locality matter more than raw parameter count.
via MarkTechPost