TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Liquid AI has introduced LFM3, a family of six models built from a shared 1.7-billion-parameter base, with text, vision, audio, multimodal and decision variants. The company says the models are designed for task-specific fine-tuning and local deployment; performance and speed figures in its announcement are company-reported benchmarks.
Liquid AI has announced LFM3, a family of six models for text, vision, audio, multimodal input and calibrated decision tasks, all branching from the LFM3-1.7B-Base model. The company presents the release as a toolkit for developers to fine-tune a model to a particular use and run it on local hardware, rather than relying on one large, general-purpose model.
Liquid AI says the family includes LFM3-1.7B for text, LFM3-VL-2B for images and video, LFM3-Audio-2B for audio input and speech output, and LFM3-Omni-2B for combined text, image, video and audio. Two decision models, d1-2B and d1-omni-2B, produce calibrated answers and are built on the vision and omni models respectively. The company describes these as six models in the 2-billion-parameter class, though their underlying configurations differ.
The shared base has 1.7 billion parameters and a 256,000-token context window. Liquid AI says it was pretrained on 22 trillion tokens, followed by a 3.5-trillion-token mid-training stage using curated data with a larger share of agent traces and tool calls. The company’s architecture combines short convolutions with sliding-window and global attention. It says this design lets the base model match the LFM2-1.2B model on short prompts per parameter and run about 25% faster with 8,000 tokens in context; these are company-reported results.
The announcement also describes new vision and audio components. Liquid AI says its 164-million-parameter vision encoder uses three times less compute than the SigLIP2 encoder used in earlier LFM vision models. For audio, the company says its causal encoder can process incoming speech as it arrives, with time to first token as low as 50 milliseconds. Its LiquidCodec audio tokenizer is reported to generate speech at about 87 times real time on one CPU thread on an Apple M4 Max, with streaming decoding.
The release targets developers who need a model adapted to specific inputs, workflows or terminology but have limited deployment resources. Liquid AI says each model can be fine-tuned on a single GPU; its text model can fit in about 1 GB at 4-bit precision. If these capabilities hold across practical deployments, teams could build local applications for tasks such as speech interfaces, visual analysis or tool-using agents without running every workload through a large cloud model.
The approach also gives developers a choice between specialized variants while keeping a common base architecture. That can matter where latency, connectivity, privacy or hardware costs influence whether an AI feature is feasible. The company’s figures are not independent evaluations, however, and real-world performance will depend on quantization, fine-tuning data, hardware and the software used to run the models.
Products Worth Considering
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
As an affiliate, we earn on qualifying purchases.
How LFM3 Builds on LFM2
Liquid AI’s previous LFM2 family used a combination of short convolutions and attention, which the company said was suited to fast on-device inference. For LFM3, the company says it redesigned the architecture through hardware-in-the-loop searches covering 14 device and software-backend combinations, ranging from phones and Raspberry Pi 5 boards to laptops, cloud processors and Qualcomm’s Hexagon NPU.
The new base uses 24 layers: 13 gated short-convolution layers, five sliding-window attention layers with a 256-token window, and six global-attention layers. The announcement also details gated attention heads, partial rotary positional embeddings and normalization at both the input and output of each block. Liquid AI says these choices are intended to improve efficiency and support long-context agent tasks. The family’s vision model adds video input, which the company identifies as a first for its LFM vision line.
“LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.”
— Liquid AI, in its Hugging Face announcement
Independent Testing Is Still Needed
The source is Liquid AI’s own product announcement, so its benchmark comparisons, speed figures and claims about relative capability are vendor-reported. The supplied material does not include independent evaluations or enough testing details to establish how the results compare across different hardware, workloads and inference settings. The claim that LFM3-1.7B outperforms LFM2.5-2.6B with fewer parameters, for example, should be read as the company’s comparison rather than a separately verified finding.
The announcement does not specify the publication date, model licensing terms, complete availability details, or the full hardware and configuration conditions for every result. It also does not give enough information to assess how reliably the models handle different languages, specialized domains or safety-sensitive tasks. Fine-tuning on one GPU is the company’s stated design target; time, memory and data requirements for a particular project remain dependent on implementation.
Model Cards and Deployment Tests
Developers can consult the LFM3 model cards and Liquid AI’s guidance for supported inference engines and agent harnesses. The company says the models run in Transformers, and its report includes a basic loading example for LFM3-1.7B. The next practical step is for users to test the relevant variant against their own data, devices and latency requirements, then compare results with alternative models under the same conditions.
Further evidence will depend on independent benchmarking and deployment reports that disclose model versions, hardware, quantization and evaluation methods. The supplied announcement does not give a timetable for additional releases or external assessments, so those remain pending rather than confirmed milestones.
Key Questions
What is LFM3?
LFM3 is Liquid AI’s family of six models for text, vision, audio, combined multimodal input and decision tasks. They branch from the LFM3-1.7B-Base model, according to the company.
Which LFM3 models can process video?
Liquid AI says LFM3-VL-2B accepts text, images and video, while LFM3-Omni-2B accepts text, images, video and audio. The d1 decision models also accept video, with d1-omni-2B handling audio as well.
Can LFM3 models be fine-tuned locally?
Liquid AI says each model is designed to be fine-tuned on a single GPU. The resources needed for a specific fine-tuning run will depend on its data, settings and software.
Are LFM3’s performance claims independently verified?
The supplied source is Liquid AI’s announcement, and its comparisons and speed figures are company-reported. The material does not include independent benchmark results that verify those claims.
Source: rss
