Liquid AI's LFM3 Squeezes Multimodal AI Into 1GB on Local Hardware
Liquid AI released LFM3 , six open-weight 2B-class models sharing one backbone across text, vision, audio, omni, and decision tasks. New tri-hybrid architecture mixes 13 gated convolutions, 5 sliding-window attention, and 6 global attention layers across 24 layers.
Takeaways
- Liquid AI released LFM3 , six open-weight 2B-class models sharing one backbone across text, vision, audio, omni, and decision tasks.
- New tri-hybrid architecture mixes 13 gated convolutions, 5 sliding-window attention, and 6 global attention layers across 24 layers.
- LFM3-Audio-2B supports 18 languages, tool calls, ~50 ms time to first token, LiquidCodec at 87x real time.
- LFM3-VL-2B averages 71.4 across 62 vision benchmarks, matching Gemma 4 12B with roughly a sixth of the parameters.
- d1-2B and d1-Omni-2B return typed structured outputs with calibrated probabilities in one forward pass.
- Available via Hugging Face, Ollama, llama.cpp and OpenVINO under the revenue-limited LFM Open License v1.0.
Liquid AI has released LFM3 , a family of open-weight models for text, vision, audio, tool use, and structured decision tasks. The six 2B-class checkpoints share an LFM3-1.7B-Base backbone and target deployment on devices, private infrastructure, and cloud servers.
Liquid says each checkpoint can be fine-tuned on a single GPU. Actual memory requirements will depend on the model, context length, batch size, precision, and training method.
| Model | Inputs | Outputs and intended use |
|---|---|---|
| LFM3-1.7B-Base | Text | General backbone for further training |
| LFM3-Audio-2B | Text and audio | Text, speech, and tool calls |
| LFM3-VL-2B | Text, images, and video | Visual question answering and analysis |
| LFM3-Omni-2B | Text, images, and audio | Joint multimodal reasoning and speech responses |
| d1-2B | Text and images | Structured answers with calibrated probabilities |
| d1-Omni-2B | Text, images, and audio | Multimodal structured decisions |
One backbone, three operators
LFM3’s 24-layer backbone combines 13 gated short-convolution layers, five sliding-window attention layers, and six global-attention layers. The local attention window covers 256 tokens, while the convolution kernel has grown from three elements in LFM2 to seven in LFM3.
Global attention can connect every token to every other token, which helps with distant dependencies but consumes more compute and memory as sequences grow. Sliding-window attention limits each token to nearby context, while short convolutions capture local patterns efficiently. LFM3 reserves global attention for layers where broad context provides the most value.
Liquid selected this mixture through automated hardware-in-the-loop search. Candidate architectures were evaluated for model quality, measured latency, and memory use across 14 processor classes. The targets included flagship phones, Apple M4 Max and AMD Ryzen AI laptops, Raspberry Pi 5, Arm and x86 cloud CPUs, GPUs, and Qualcomm’s Hexagon NPU.
Liquid reports that LFM3-1.7B outperforms LFM2.5-2.6B in its evaluations while using about 35% fewer parameters. Quantized to 4-bit precision, the weights occupy roughly 1 GB. Total runtime memory will be higher because inference also requires a key-value cache, activations, buffers, and application overhead.
The base model supports a context window of up to 256,000 tokens, aimed at long documents and agent sessions. Maximum context capacity does not establish accuracy across that full range, so long-context quality and memory consumption require testing in the intended runtime.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
