Liquid AI's LFM3 Squeezes Six Multimodal Models Into 2B Parameters
Liquid AI's LFM3 family puts six specialized 2B models on one tri-hybrid backbone, with vision, audio, and decision variants that fine-tune on a single GPU.
- Liquid AI released LFM3, six open-weight 2B-class models sharing one tri-hybrid backbone.
- Lineup spans text, vision-with-video, audio, omni, and typed-output decision models.
- LFM3-VL-2B averages 71.4 on 62 vision benchmarks, matching Gemma 4 12B at a sixth the size.
- Architecture mixes 13 short convolutions, 5 sliding-window attention, 6 global attention layers.
- Text model fits in roughly 1 GB at 4-bit precision with a 256K context window.
- Every model fine-tunes on a single GPU; weights live on Hugging Face under LFM Open License v1.0.
Liquid AI has released LFM3, a family of six open-weight models for text, vision, audio, multimodal input, and structured decisions. The models range from 1.7 billion to roughly 2 billion parameters and use a 24-layer architecture that combines short convolutions, local attention, and global attention to reduce memory use during inference.
The release targets workloads that need low latency, local execution, or extensive fine-tuning without the cost of serving a larger general-purpose model. Its flagship vision model, LFM3-VL-2B, contains 1.92 billion parameters and accepts images and video alongside text.
Six branches from one base
The family derives from a shared pretrained foundation, with each branch adapted for a specific input type or output format.
| Model | Primary use | Key details |
|---|---|---|
| LFM3-1.7B | Text agents | 256K-token context window |
| LFM3-VL-2B | Vision and language | Image and video input; 128K-token context |
| LFM3-Audio-2B | Speech applications | Voice chat, speech recognition, speech synthesis, and tool calling in 18 languages |
| LFM3-Omni-2B | Multimodal applications | Combines vision, audio, and language |
| d1-2B | Structured decisions | Typed outputs and calibrated probabilities |
| d1-Omni-2B | Multimodal decisions | Structured predictions from multimodal input |
Liquid publishes GGUF, MLX, ONNX, FP8, and NVFP4 variants, along with hosted access through its playground and API. The weights use the LFM Open License v1.0, which permits free commercial use for organizations with less than $10 million in annual revenue. Teams approaching that threshold should review the custom license before deployment.
A cache-saving layer mix
Full self-attention compares every token with every other token during prompt processing, causing compute and attention-memory requirements to rise quadratically with sequence length. During generation, transformers also retain key-value states for earlier tokens, so the KV cache grows with the context window and number of attention layers.
LFM3’s 24-layer language backbone limits those costs by assigning different jobs to three layer types:
- 13 gated short-convolution layers: Use a kernel size of 7, maintain constant-size recurrent state, and add nothing to the KV cache.
- 5 sliding-window attention layers: Attend to a 256-token window, placing a fixed bound on their cache use.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
