LiquidAI releases six LFM3 models built for specialized, on-device AI

LiquidAI is promising specialized AI without a large general-purpose model. On October 27, 2026, it released LFM3, six models sharing one foundation for text, vision, audio and decision tasks. The company says each can be adapted on a single GPU and deployed on-device; its performance comparisons remain company-reported results.

All six models have downloadable weights on Hugging Face. Rather than asking developers to use every capability at once, LiquidAI offers separate text, vision and audio models, a combined model, and two decision variants. They all branch from LFM3-1.7B-Base and sit around the two-billion-parameter size range.

Choose the inputs before choosing the model

• LFM3-1.7B handles text. LiquidAI positions it for local agents, tool calling, structured responses and further training on specialized data.
• LFM3-VL-2B takes text, images and video and produces text. It is LiquidAI’s first vision model to accept video.
• LFM3-Audio-2B takes text or audio and returns text, speech or both. Its stated uses include voice chat, transcription, speech generation and translation across 18 languages.
• LFM3-Omni-2B combines text, images, video and audio as inputs, with text and speech outputs. LiquidAI says it can answer spoken questions about images and process a video’s pictures and soundtrack together.
• The decision models, d1-2B and d1-omni-2B, build on the vision and combined models respectively. LiquidAI lists their outputs as calibrated answers; the omni version also accepts audio.

Omni is a first for the company. LiquidAI presents it as an alternative to assembling separate systems for speech recognition, visual understanding, language and speech generation. It also says the model can turn what it sees and hears into tool calls, connecting mixed-media input to actions in software.

A small backbone with a long text window

The text model has a 256K-token context window—the amount of material it can work with at once. LiquidAI says it fits in about 1 GB at four-bit precision, a compressed numerical format. That size claim applies to the text model, not to the entire multimodal lineup.

Underneath, LFM3 mixes three kinds of processing across 24 layers: short convolutions, attention over a nearby window, and global attention. The design combines local processing with layers that connect information across the input. LiquidAI selected the architecture using a search covering 14 device-and-software combinations, including phones, Raspberry Pi 5, laptops and cloud hardware.

Training also reflects the focus on agents. The base model was trained on 22 trillion tokens. The final 3.5 trillion-token stage used curated data with more records of agent actions and tool calls, while extending the context window to 256K tokens.

The gains LiquidAI reports—and the route to running it

LiquidAI reports that LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters. Its clearest cited agent comparison is WebArena-Verified, where the new model scores 38.42 against its predecessor’s 3.02. Those figures describe the company’s benchmark evaluation, rather than a guarantee about a developer’s particular workflow.

The vision model follows the same smaller-model argument. LiquidAI reports an average score of 71.4 across 62 vision benchmarks, up from 57.6 for LFM2.5-VL-3B. Parameter count falls from 3.1 billion to 1.9 billion. Its new vision encoder also compresses each pair of sampled video frames into one temporal slice, halving that sequence’s length.

For developers, the launch includes model-loading examples using Hugging Face’s Transformers library. LiquidAI also links fine-tuning notebooks and the base checkpoint for teams doing their own additional training. The stated path is to choose the inputs an application needs, adapt the corresponding model on one GPU, and deploy it where the application runs.