🔍 Read the full analysis: Steps For Creating Specialized Multimodal Models With LFM3 on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Liquid AI has introduced LFM3, a family of models based on a shared 1.7-billion-parameter foundation and adapted for text, vision, audio, multimodal inputs and decision tasks. The company says they are designed for single-GPU fine-tuning and deployment on edge devices, but the reported performance figures have not been independently verified in the supplied material.
Liquid AI has introduced LFM3, a family of models built around a shared 1.7-billion-parameter base and adapted for text, images, video, audio and decision tasks. The company says developers can fine-tune the models on one GPU and deploy them on edge devices or other infrastructure, positioning the release for teams seeking specialized multimodal systems; its performance and efficiency figures remain company-reported.
The family includes LFM3-1.7B for text, LFM3-VL-2B for text, images and video, LFM3-Audio-2B for text and audio, and LFM3-Omni-2B for text, image, video and audio inputs. Liquid AI also introduced two decision models, d1-2B and d1-omni-2B, which it describes as producing calibrated answers. Each model branches from LFM3-1.7B-Base, with additional modality-specific components where needed.
Liquid AI describes the base as a 24-layer tri-hybrid architecture combining short convolutions, sliding-window attention and global attention. The company says LFM3-1.7B was trained on 22 trillion tokens, including a final 3.5-trillion-token stage using curated data with more agent traces and tool calls. It reports that the model outperforms its LFM2.5-2.6B predecessor with 35% fewer parameters, and runs about 25% faster than LFM2-1.7B at an 8,000-token context. These comparisons are attributed to Liquid AI; the supplied material does not include independent evaluations or full testing protocols.
The text model is listed with a 256,000-token context window, and Liquid AI says it fits in about 1 GB at 4-bit precision. For vision, the company reports that a 164-million-parameter encoder matches the accuracy of Qwen3.5-2B’s vision encoder while using 2.2 times fewer floating-point operations in a comparison using the same LFM2.5-230M language backbone. Liquid AI also says its audio system can process speech as a person is still speaking. The supplied source cites a LiquidCodec speech-generation rate of about 87 times real time on one CPU thread on an Apple M4 Max. Those figures describe specific company-reported tests, not independently established results across hardware and workloads.
Why Smaller Models Could Matter
LFM3’s design targets a practical choice for developers: select a model for the inputs and task they need, rather than rely on a single general-purpose system for every use. If the company’s size, speed and fine-tuning claims hold in real deployments, smaller specialized models could make it easier to run some AI tasks close to where data is produced, including on local devices or private infrastructure. That could be relevant for applications where latency, connectivity or control over data flows affects system design.
The vision model’s video input and the audio model’s streaming capability also point to possible uses in interactive assistants, robotics and other systems that respond to changing inputs. These are potential applications, not results demonstrated by the announcement. It does not establish that LFM3 will lower operating costs, improve privacy or outperform larger hosted models in production. Those outcomes depend on task quality, hardware, integration and the amount of fine-tuning required.
Popular right now
iRobot Roomba Max 715 Pet Robot Vacuum with AutoEmpty Dock
Dyson Car+Boat Cordless Handheld Vacuum
Breville Bambino Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
Liquid AI presents LFM3 as the next generation of its Liquid Foundation Models, following the LFM2 family. The company says its earlier models combined short convolutions with attention to support efficient operation on devices. For LFM3, it reports that a hardware-in-the-loop architecture search covered 14 combinations of devices and computing backends, ranging from phones and a Raspberry Pi 5 to laptops, cloud CPUs and GPUs, and Qualcomm’s Hexagon NPU.
The announced design changes include convolution layers with a kernel size of seven, compared with three in LFM2, and head-wise gates that can scale individual attention heads. Liquid AI also describes partial rotary position embeddings in global attention layers and normalization at the input and output of each block. It says these choices aim to improve capability per parameter and support longer-context inference. The models are available through Transformers according to the supplied material, which directs developers to model cards for inference engines and setup information.
“LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.”
— Liquid AI
Independent Testing Still Needed
The supplied material provides no independent benchmark results, complete evaluation protocols or detailed comparisons across deployment conditions. The reported results come from Liquid AI, and one vision-benchmark description in the source is incomplete. It is also not clear how much training data, time or GPU memory a particular fine-tuning job will require, or how the models perform on production workloads.
Model size at 4-bit precision and a long context window do not by themselves establish end-to-end latency, output quality or operating cost. The supplied information also does not specify licensing terms, whether all model weights are available, or release dates for each model and deployment option. Developers will need those details to judge how readily the family can be used in a specific application.
Model Cards and Deployment Checks
Liquid AI points developers to the LFM3 model cards and an agent-harness guide for supported inference engines and setup instructions. The company says the models run in Transformers, with an example available for loading the text model. The next useful test will be whether teams can reproduce the reported results and fine-tune the models for their own tasks on the stated hardware.
Further assessment will depend on full benchmark methods, access and licensing details, and measurements of memory use, latency and task quality under comparable conditions. Until those are available, LFM3’s claimed efficiency and suitability for edge deployments remain proposals supported by company-reported figures rather than independently confirmed outcomes.
Key Questions
What is LFM3?
LFM3 is a model family from Liquid AI built from a shared 1.7-billion-parameter base, with versions adapted for text, vision, audio, multimodal inputs and decision tasks.
Which inputs do the LFM3 models support?
The announced range covers text, images, video and audio, in different combinations. LFM3-VL-2B handles text, images and video; LFM3-Audio-2B handles text and audio; and LFM3-Omni-2B combines all four input types.
Can developers fine-tune LFM3 on one GPU?
Liquid AI says the models are designed for single-GPU fine-tuning. The supplied information does not specify the data, memory, time or GPU requirements for individual fine-tuning jobs.
Have LFM3’s performance claims been independently verified?
Not in the supplied material. The speed, accuracy and efficiency figures are reported by Liquid AI, and the source does not provide independent evaluations or full benchmark protocols.
When will all LFM3 models be available, and under what license?
The supplied information does not give release dates for each model or licensing terms. Liquid AI points developers to model cards and setup guidance for implementation details.
Primary source: Hugging Face · via ThorstenMeyerAI.com
