The Pulse
Liquid AI Releases Six Models on One Small-Model Backbone
Liquid AI introduced LFM3, a family of six small models covering text, vision, video, audio and structured decisions. The models share a base architecture and are presented as adaptable for different data and deployment settings.
AI.info Team · Written with AI · fact-checked ·
“Liquid AI's LFM3 family combines strong multimodal capabilities with a compact footprint that is well suited for on-device AI experiences.”
Josh Hort, Senior Director and Head of ISV Enabling at AMD
Liquid AI published LFM3 on October 27, presenting six small models for text, images, video, audio and structured decisions. The models share a base architecture, giving developers a common starting point for choosing a modality, adapting the model to their own data and running it where their application operates. Liquid AI says the open-weight models are available on Hugging Face.
The family spans LFM3-1.7B for text; LFM3-VL-2B for vision; LFM3-Audio-2B; LFM3-Omni-2B; and two decision models, d1-2B and d1-omni-2B. The vision model accepts video as well as images, while Omni combines text, image, video and audio input with text and speech output. The decision variants answer structured questions about an input rather than generating a conventional free-form response.
One base, six specialized models
Liquid AI says all six models are trained from LFM3-1.7B-Base, with modality-specific encoders, decoders or post-training added where needed. That shared foundation is intended to let teams move between related applications without starting from unrelated model architectures. The company positions fine-tuning as a way to adapt a model to a particular dataset, vocabulary or deployment setting.
The base model uses 24 layers combining three approaches: gated short convolutions, sliding-window attention and global attention. Liquid AI describes the mix as a tri-hybrid architecture developed through hardware-in-the-loop searches across 14 device and software-backend combinations, including phones, a Raspberry Pi 5, laptops, cloud processors and Qualcomm’s Hexagon NPU. The company says it designed the changes to improve capability per parameter and inference over long contexts.
LFM3-1.7B also uses a 256,000-token context window, according to the release. Liquid AI says a four-bit version of the text model occupies about 1 GB, a size aimed at local use on devices with limited memory. The company says the model can be fine-tuned on one GPU, though the post does not set out a minimum memory requirement or a single hardware configuration for that claim.
Liquid AI’s reported gains vary by task
Liquid AI says LFM3-1.7B outperforms its LFM2.5-2.6B predecessor while using 35% fewer parameters. On WebArena-Verified, the company reports a score of 38.42 for LFM3-1.7B against 3.02 for LFM2.5-2.6B. Those figures are the company’s published results; the release does not present an independent assessment.
For vision, Liquid AI reports an average score of 71.4 across 62 benchmarks for LFM3-VL-2B, compared with 57.6 for its earlier LFM2.5-VL-3B. The post says the new vision model has 1.9 billion parameters, compared with 3.1 billion for its predecessor. The model’s video input marks a first for Liquid AI’s LFM vision line.
The release also describes new in-house vision and audio encoders. Liquid AI says its vision encoder uses three times less compute than the SigLIP2 encoder used in earlier LFM vision-language models. For audio generation, the company says its LiquidCodec system can produce speech faster than real time; its audio model supports speech recognition, speech synthesis, translation across 18 languages and tool calls.
Omni and decision models target different workflows
LFM3-Omni-2B combines audio, image and video understanding, and can respond in text or speech. Liquid AI says it can answer spoken questions about images, process audio and video together, and turn multimodal inputs into tool calls. The company also reports that Omni outperforms models up to 10.7 billion parameters on three audio-visual video benchmarks: Daily-Omni, WorldSense and AV-SpeakerBench.
The d1 models take a different route: they return structured answers, such as classifications, choices or scores, with calibrated probabilities. Liquid AI says d1-2B scores 75.5 on ImajevBench, versus 64.0 for the previous d1-3B, and that the audio-capable d1-omni-2B retains nearly all of d1-2B’s accuracy when questions arrive by voice. The post does not provide a full results table in its main summary, so those claims are best read alongside the model-specific evaluation details.
All six models are available through Hugging Face, and Liquid AI links notebooks for fine-tuning text, vision, audio and Omni versions. The release’s concrete proposition is a shared starting model rather than a single all-purpose system: developers select a specialist variant, adapt it, and deploy it on their own hardware or infrastructure.