Liquid AI Launches LFM3 Open-Weight Multimodal Models for Edge AI
Liquid AI launches six open-weight LFM3 edge models spanning text, vision, audio and decisions, with single-GPU tuning and a 256K context window, on devices.
Summary
On October 27, 2026, Liquid AI launched six open-weight, roughly 2B-parameter LFM3 models from one 1.7B backbone: LFM3-1.7B for text, LFM3-VL-2B for images and video, LFM3-Audio-2B, LFM3-Omni-2B, plus d1-2B and d1-omni-2B for calibrated decisions. Every model supports single-GPU fine-tuning and Transformers deployment, while the 4-bit text model occupies about 1 GB and supports 256K tokens. The 24-layer tri-hybrid backbone combines 13 gated convolutions, five 256-token sliding-window attention layers and six global attention layers. It was trained on 22T tokens, including 3.5T curated mid-training tokens, using a 128K vocabulary and architecture searches across 14 hardware and backend combinations. At 8K context, LFM3-1.7B is about 25% faster per parameter than LFM2-1.2B.
LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters and scores 38.42 versus 3.02 on WebArena-Verified. LFM3-VL-2B raises the 62-benchmark vision average from 57.6 to 71.4 despite shrinking from 3.1B to 1.9B parameters, matching Gemma 4 12B; its 164M encoder uses three times less compute than SigLIP2 and 2.2 times fewer FLOPs than Qwen3.5-2B’s encoder. LFM3-Audio-2B supports 18 languages, 50 ms first-token latency and LiquidCodec speech generation at about 87 times real time on one Apple M4 Max CPU thread, improving UTMOSv2 from 3.22 to 3.83. LFM3-Omni-2B accepts text, images, video and audio, outperforming models up to 10.7B on three audiovisual benchmarks. d1-2B replaces d1-3B with one-third fewer parameters and scores 75.5 versus 64.0 on ImajevBench. A robotics adaptation reaches 96.2% across LIBERO, 45.7% on RoboCasa-GR1 and under 50 ms action generation on Jetson AGX Orin. All models and fine-tuning notebooks are available on Hugging Face.
Positives
- 35% fewer parameters lets LFM3-1.7B outperform LFM2.5-2.6B while fitting into about 1 GB at 4-bit precision.
- 71.4 across 62 vision benchmarks puts the 1.9B-parameter LFM3-VL-2B level with the much larger Gemma 4 12B.
- 50 ms first-token latency and 87 times real-time speech decoding support responsive voice applications on edge hardware.
- 256K tokens, single-GPU fine-tuning and open weights make customized local agents practical without frontier-scale infrastructure.
- 96.2% average LIBERO success demonstrates compact LFM3-VL-2B adaptations can control robots in real time on embedded hardware.
Risks & concerns
- About 1 GB of memory requires 4-bit precision, so deployments at higher precision need more capacity.
- Single-GPU fine-tuning lowers infrastructure requirements but still requires dedicated accelerator hardware.
- 18-language audio support limits LFM3-Audio-2B for applications requiring broader language coverage.
- Nearly all accuracy is the only quantified description of d1-omni-2B voice performance, leaving its exact degradation unspecified.