Introducing LFM3: Every Modality for Specialized Intelligence

October 27, 2026 — Liquid AI is proud to unveil LFM3, the latest generation of Liquid Foundation Models. This family is specifically engineered to deliver specialized intelligence across a diverse array of modalities.

Unlike previous iterations, the LFM3 suite is natively multimodal, unifying text, vision, audio, omni-capabilities, and decision-making logic onto a single, shared architectural backbone.

🚀 Quick Overview

The initial rollout focuses on models in the 2B parameter class. By leveraging a novel hybrid architecture and a massive pre-training dataset of $22 \times 10^{12}$ (22T) tokens, LFM3 achieves significantly higher expressivity than LFM2 while remaining lean and efficient.

Because these models are small and fast, they serve as the perfect foundation for specialized AI:

• Local Fine-tuning: Adaptable on private data.
• Universal Deployment: Capable of running on virtually any hardware.
• Expert Performance: Sufficient capacity to master niche, complex tasks.

Availability Checklist

• Open-weights on Hugging Face
• Interactive Playground
• Liquid AI API

---

🌟 Key Highlights

---

🛠️ A New Architecture for Enhanced Expressivity

While LFM2 proved that combining short convolutions with attention enabled fast on-device performance, LFM3 evolves this further. Using an automated hardware-in-the-loop architecture search, Liquid AI redesigned the system for agentic use cases, prioritizing high capability-per-parameter and rapid inference across extensive contexts.

Hardware-in-the-Loop Testing

The search space was expanded from a few devices to 14 distinct combinations, including:

• Mobile: Samsung Galaxy S26 Ultra, OnePlus 15.
• SBC: Raspberry Pi 5.
• Laptops: Apple M4 Max, AMD Ryzen AI.
• Cloud/Compute: Arm and x86 CPUs, GPUs, and the Qualcomm Hexagon NPU.

The Tri-Hybrid Operator Mix

To optimize the speed-to-quality ratio for long contexts, LFM3 implements Sliding-Window Attention (SWA).

Technical Specifications:

1. Convolutions: Kernel size increased to $7$ (up from $3$ in LFM2).
2. Sigmoid Gating: Each head's output is scaled via a sigmoid gate, allowing the model to "switch off" unnecessary heads. This reduces activation outliers, which is critical for $\text{quantization}$ and $\text{training stability}$.
3. Global Attention: Only the high-frequency half of each head is rotated; the low-frequency half remains static to ensure content consistency regardless of distance.
4. Normalization: Both input and output of each block are normalized to maintain well-scaled activations.

Training & Data

LFM3 utilizes the same $128\text{K}$ vocabulary as LFM2.5-8B-A1B for better non-Latin script efficiency. The training mix was expanded to include:

• More scientific literature and repository-level code.
• Diverse knowledge sources and increased mathematical data.
• Mid-training phase: The final $3.5\text{T}$ tokens consist of high-quality, curated data featuring agentic traces and tool-calling examples.

---

🌐 A Multimodal Family for Every Need

The LFM3 release consists of six distinct models, all branching from the LFM3-1.7B-Base core.

Model Lineup:

• Input Modalities: Text, Vision, Audio, and Omni.
• Specialized Logic: Two Decision models (derived from the Vision and Omni variants).

Instead of relying on a single, bloated general-purpose model, developers can select the specific modality required, customize it, and deploy it directly to the edge.

---

🎙️ Industry Perspectives

"As AI increasingly moves from the cloud to the client, developers need models that can be deployed efficiently across the broad spectrum of AI PCs. Liquid AI's LFM3 family combines strong multimodal capabilities with a compact footprint... we're excited to see LFM3 expand what's possible for specialized AI applications running locally on systems with AMD Ryzen™ AI processors." — Josh Hort, Senior Director Head of ISV Enabling at AMD

"Across industries, we’re seeing edge AI evolve from running dedicated detection and perception models to supporting multiple concurrent AI workloads... LFM3 brings text, vision, and audio together in a unified architecture designed for efficient deployment... Developers can already run Liquid AI’s LFM2.5-VL models on Ambarella silicon through the Cooper Model Garden, and we have begun enabling LFM3 on the same platform." — Malhar Palkar, Senior Director for AI at Ambarella