Liquid AI Unveils LFM3 On-Device Tri-Hybrid Foundation Models
Liquid AI introduced the LFM3 family of 2B-class models built on a tri-hybrid architecture, supporting multimodal edge workloads and single-GPU fine-tuning.
On October 28, 2026, Liquid AI officially introduced 'LFM3', its next-generation family of foundation models built on a novel tri-hybrid architecture. The release features six 2B-class models covering text, vision, audio, omni, and decision-making tasks.
Image source: @liquidai
Image Credit: Liquid AI
Designed via automated hardware-in-the-loop architecture search targeting 14 processor classes, LFM3 targets capability per parameter and fast inference over long contexts for local execution across edge devices, PCs, and accelerators.
Tri-Hybrid Architecture and 2B-Class Model Suite
The core innovation behind LFM3 is its tri-hybrid architecture, which combines gated short convolutions, sliding-window attention, and global attention.
- LFM3-1.7B: A text model supporting a 256K-token context window, with weights fitting in approximately 1GB at 4-bit precision and a 1.6GB KV cache footprint at 128K tokens.
- LFM3-VL-2B: A 1.9B vision-language model with video input and spatial reasoning capabilities.
- LFM3-Audio-2B: A model supporting voice chat, tool calls, and reasoning across 18 languages.
- LFM3-Omni-2B: A multimodal model combining vision, audio, and language.
- d1-2B and d1-Omni-2B: Decision models offering typed outputs and calibrated probabilities in a single forward pass.
Edge Deployment and Single-GPU Customization
LFM3 models are built for specialization, allowing developers to fine-tune them on a single GPU using domain-specific data.
- Hardware Targeting: Optimized across smartphones (Galaxy S26 Ultra, OnePlus 15), laptops (Apple M4 Max, AMD Ryzen AI), Raspberry Pi 5, ARM/x86 cloud CPUs, and the Qualcomm Hexagon NPU.
- Evaluation Results: According to Liquid AI's evaluations, LFM3-1.7B outperformed Qwen3.5-9B on WebArena-Verified and AutomationBench, while LFM3-VL-2B, fine-tuned on approximately 3,000 public WTBs2025 wind-turbine photos, reached 67% defect-classification accuracy on approximately 1,400 held-out photos.
Sources
- Liquid AI Official X (@liquidai): 2026-10-28 LFM3 Announcement Thread
- Liquid AI Official Blog: LFM3 Release Overview