Liquid AI's LFM3 Beats 9B Rivals Using a Tiny 1.7B On-Device Model
Liquid AI launched LFM3, a tri-hybrid family of six open-weight 2B models spanning text, vision, audio, omni, and decision tasks, all fine-tunable on a single GPU.
- Liquid AI released LFM3, six open-weight 2B models from one shared backbone across text, vision, audio, omni, decision.
- New tri-hybrid architecture: gated short convolutions, sliding-window attention, and global attention, searched across 14 hardware targets.
- LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters and tops Qwen3.5-9B on agentic benchmarks.
- LFM3-VL-2B averages 71.4 across 62 vision benchmarks, matching Gemma 4 12B at a sixth the size.
- d1-2B returns typed outputs and calibrated probabilities in one forward pass, 5.7 ms on an H100.
- Open weights on Hugging Face, 1 GB at 4-bit, 256K context, fine-tune on a single GPU.
Liquid AI has released LFM3, a family of six models derived from the same 1.7-billion-parameter backbone and sized for local deployment. The LFM3 release spans text, vision, audio, multimodal generation, and structured decision models. The vision-capable d1-2B weights drew particular attention because the model returns typed answers and probabilities in one forward pass.
The family targets developers who need lower latency, local data processing, or predictable inference costs. Liquid says each model can be fine-tuned on a single GPU and deployed on phones, laptops, private servers, or cloud infrastructure. Developers select a specialized model instead of serving the entire family.
Six models from one backbone
Liquid trained the shared LFM3-1.7B-Base backbone on 22 trillion tokens, then adapted it for six model variants:
| Model | Size | Inputs and outputs | Primary use |
|---|---|---|---|
| LFM3-1.7B | 1.7B | Text | Local assistants and agent systems with a 256K-token context window |
| LFM3-VL-2B | 1.9B | Images, video, and text | Visual and spatial reasoning |
| LFM3-Audio-2B | About 2B | Audio and text | Voice chat, reasoning, and tool calls in 18 languages |
| LFM3-Omni-2B | 2.2B | Vision, audio, and language | Multimodal responses in text, speech, or both |
| d1-2B | About 2B | Text, JSON, images, and video | Typed decisions, extraction, classification, and localization |
| d1-Omni-2B | About 2B | Multimodal state | Structured decisions across multiple modalities |
Three layer types trim inference costs
Liquid’s “tri-hybrid” architecture combines gated short convolutions, sliding-window attention, and global attention. Short convolutions process nearby token patterns cheaply. Sliding-window layers extend the working range while limiting how many tokens each position can inspect. Global attention lets tokens interact across the full context, carrying the highest memory and compute cost.
The 24-layer d1-2B stack contains 13 gated short-convolution layers, five sliding-window attention layers, and six global-attention layers. Liquid selected those ratios through a hardware-in-the-loop search that scored model quality, measured latency, and memory use across 14 processor classes. Targets included the Galaxy S26 Ultra, OnePlus 15, Apple M4 Max, Ryzen AI laptops, Raspberry Pi 5, and Qualcomm Hexagon NPU.
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
