Liquid AI Releases LFM3 Multimodal 2B-Class Model Family

For audio-centric applications, the LFM3-Audio-2B introduces a custom in-house audio encoder and codec. It supports 18 languages and enables low-latency voice chat with streaming replies and integrated tool calling. Finally, the d1-2B and d1-Omni-2B decision models are designed for high-speed tasks like routing, triage, and verification, providing calibrated probabilities in a single pass within milliseconds.

• LFM3-VL-2B: Video input support with step-by-step thinking
• LFM3-Audio-2B: 18 languages, streaming voice chat, and tool calling
• LFM3-Omni-2B: Combined vision, audio, and language in one model
• d1-series: Decision models for rapid triage and routing with calibrated probabilities

Performance & Benchmarks

LFM3 redefines the Pareto frontier of device intelligence. The LFM3-1.7B text model demonstrates incredible efficiency, outperforming the previous LFM2.5-2.6B model despite having 35% fewer parameters. For developers working with constrained memory, the text model features a massive 256K context window and can fit into approximately 1 GB of VRAM at 4-bit precision.

In multimodal benchmarks, the results are even more striking. The LFM3-VL-2B vision model ranks at least 7.4 points ahead of all other models in its size class. Remarkably, it matches the performance of Gemma 4 12B—a model over six times its size—and sits within just 2 points of the GPT-5.6 Luna. These evaluations were conducted using standardized vision-language reasoning benchmarks to ensure parity with industry leaders.

• LFM3-1.7B: 256K context window, ~1GB footprint at 4-bit
• VL Benchmarks: +7.4 points over similar-sized rivals
• Comparison: Matches Gemma 4 12B; nears GPT-5.6 Luna performance
• Efficiency: Ranked #1 on the device-intelligence Pareto frontier (pipette.liquid.ai)

Developer Implementation & Use Cases

LFM3 is designed for the modern developer workflow. Every model in the family is fine-tunable on a single GPU, significantly lowering the barrier to entry for creating specialized agents. Whether you are building a RAG (Retrieval-Augmented Generation) system that requires a massive context window or a real-time voice assistant, LFM3 provides the necessary primitives.

Key use cases include: coding assistants that run locally, real-time video analysis for security or education, high-speed decision engines for app automation, and multilingual voice interfaces for mobile applications. The open-weight nature under the LFM 1.0 license allows for deep integration into private enterprise environments.

• Coding & Reasoning: High-performance text backbone
• Agents: Optimized for decision-making and tool calling
• RAG: 256K context window for large document processing
• Local Deployment: Optimized for mobile and edge NPU/CPUs

Getting Started

Developers can access the LFM3 family immediately through several channels. The open weights are available on Hugging Face for local deployment and fine-tuning. For those preferring a managed environment, the Liquid AI Playground offers a sandbox for testing different modalities, while the Liquid API provides scalable access for production-grade applications.

• Hugging Face: Open weights for local deployment
• Liquid AI Playground: Interactive testing environment
• Liquid API: Scalable cloud-based inference
• Fine-tuning: Supported on a single GPU

Sources