Liquid AI LFM3 release notes

Overview of LFM3

Liquid AI's LFM3 is a new generation of foundation models built to serve as a starting point for specialized intelligence. The family consists of six 2B-parameter-class models—covering text, vision, audio, omni, and decision-making—all sharing a common 1.7B-parameter base architecture. These models are designed for edge deployment, enabling users to fine-tune them on a single GPU for specific domains, terminology, or deployment environments.

Architectural Improvements

The LFM3-1.7B-Base model features a tri-hybrid architecture optimized for agentic use cases, which require high capability per parameter and efficient long-context inference. The architecture was developed using hardware-in-the-loop search across 14 device and backend combinations, including mobile CPUs, GPUs, and NPUs.

Key technical features of the LFM3-1.7B-Base include:

  • Tri-hybrid layer mix: A combination of 13 gated short convolutions, 5 sliding-window attention (SWA) layers with a 256-token window, and 6 global attention layers.
  • Head-wise gated attention: A sigmoid gate allows individual heads to deactivate when they have no relevant information to add, helping to suppress activation outliers and improve quantization.
  • Partial RoPE: Global attention layers apply rotation only to the high-frequency half of each head, while SWA layers maintain full Rotary Positional Embeddings (RoPE).
  • Sandwich norm: Normalization is applied to both the input and output of each block to maintain stable activations.

These improvements allow LFM3-1.7B to outperform the previous LFM2.5-2.6B while using 35% fewer parameters. The model supports a 256K context window.

Multimodal Capabilities

LFM3 offers specialized models built on the shared 1.7B backbone, incorporating new in-house encoders for vision and audio:

  • Vision Encoder: Combines convolution and attention, requiring 3x less compute than the SigLIP2 encoder used in previous versions. It supports video input by embedding frame pairs into single temporal slices.
  • Audio Encoder and LiquidCodec: The audio encoder is fully causal, enabling low time-to-first-token (as low as 50ms). LiquidCodec provides streaming speech generation at approximately 87x real-time speed on a single CPU thread.
  • Omni Model: The LFM3-Omni-2B integrates these encoders to handle text, image, video, and audio inputs, with text and speech output, replacing the need for separate models for ASR, vision, and TTS.
  • Decision Models (d1): The d1-2B and d1-omni-2B models are designed for structured decision-making. They process inputs and answer multiple questions in a single forward pass, providing calibrated probabilities for tasks like classification or extraction.

Deployment and Fine-Tuning

LFM3 is designed for on-device deployment, with the text model occupying approximately 1 GB at 4-bit precision. Liquid AI provides a suite of fine-tuning notebooks covering various methods (SFT, DPO, GRPO, CPT) to adapt the models for tasks such as structured JSON extraction, tool calling, and robotics.

For robotics applications, the LFM3-VL-2B model has been demonstrated in orchestration, navigation, and manipulation tasks. When trained across 15 embodiments, it achieved high success rates on benchmarks like LIBERO and RoboCasa-GR1, with real-time action generation on embedded hardware like the Jetson AGX Orin.

Sources