🔍 Read the full analysis: How LFM3 Supports Specialized Multimodal AI Development on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Liquid AI has introduced LFM3, a family of models built from a shared 1.7-billion-parameter base and adapted for text, vision, audio, multimodal inputs and decision tasks. The company says the models are designed for fine-tuning on one GPU and deployment on edge devices, but its performance figures have not been independently verified in the supplied material.
Liquid AI has introduced LFM3, a family of models built from a shared 1.7-billion-parameter base and adapted for text, vision, audio, multimodal inputs and decision tasks, as detailed in the original analysis of LFM3. The company says developers can fine-tune the models on one GPU and deploy them on edge devices or other infrastructure, positioning the family for specialized applications that may not need a large, general-purpose hosted model.
The lineup 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 combined text, image, video and audio inputs. Liquid AI also lists two decision models, d1-2B and d1-omni-2B, which it describes as producing calibrated answers. The models branch from LFM3-1.7B-Base, with additional encoders or decoders for different modalities.
Liquid AI says the base model uses a 24-layer tri-hybrid design that combines short convolutions, sliding-window attention and global attention. The company reports that LFM3-1.7B outperforms its LFM2.5-2.6B predecessor while using 35% fewer parameters, and is about 25% faster than LFM2-1.7B at an 8,000-token context. It gives the text model a 256,000-token context window. These comparisons and figures are company-reported; the supplied source does not include independent test results or full evaluation protocols.
According to Liquid AI, LFM3-1.7B-Base was trained on 22 trillion tokens, including a final 3.5-trillion-token stage using curated data with more agent traces and tool calls. The company says the text model fits in about 1 GB at 4-bit precision. It also reports that its 164-million-parameter vision 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.
LFM3’s design gives developers a set of modality-specific options from one underlying model family, rather than requiring every task to use the same general-purpose system. If Liquid AI’s claims about model size, speed and fine-tuning hold in practical deployments, smaller teams could have more options for adapting AI to particular tasks and running it close to where data is produced.
That approach may matter for applications with low-latency or local-processing needs, including interactive assistants, robotics and devices that handle image, video or audio inputs. Streaming audio support, for example, could let an application begin processing before a speaker finishes. These are potential uses implied by the features, not demonstrated outcomes in the supplied announcement. It does not establish real-world performance, cost reductions or improved privacy.
The announcement’s claims also make independent testing important. A smaller parameter count or longer context window does not, on its own, show how well a model performs on a specific task, how much memory it uses in an end-to-end deployment, or what it costs to run. Those results will depend on the hardware, software and fine-tuning required for each application.
Products Worth Considering
Sonos Era 100 SL Compact Microphone-Free Speaker, Black
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
As an affiliate, we earn on qualifying purchases.
How LFM3 Builds on LFM2
Liquid AI presents LFM3 as the next generation of its Liquid Foundation Models, following the LFM2 family. The company says earlier LFM2 models combined short convolutions with attention to support efficient operation on devices. For LFM3, it says an architecture search tested 14 combinations of devices and computing backends, including phones, a Raspberry Pi 5, laptops, cloud CPUs and GPUs, and Qualcomm’s Hexagon NPU.
The new base model makes several architectural changes, according to the company. Its convolution layers use a kernel size of seven, compared with three in LFM2, while head-wise gates can scale individual attention heads. Liquid AI also describes partial rotary position embeddings in global attention layers and normalization at both the input and output of each block. The company says these choices aim to improve capability per parameter and support longer-context inference; the supplied source does not provide independent evidence for those aims.
The source says the models are available through Transformers and directs developers to model cards for inference engines and setup details. It also points to an agent-harness guide. Access, licensing and deployment options are not fully specified in the supplied material.
““LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.””
— Liquid AI
Benchmark Methods Still Needed
The performance and efficiency figures in the announcement are claims from Liquid AI. The supplied material does not include independent evaluations, full benchmark methods or enough detail to compare results across different hardware and deployment conditions. It begins to describe vision-model benchmark results but is cut off before providing the full comparison.
It is also unclear how much fine-tuning data, time and memory specific applications will require, or how the models perform under production workloads. A reported 256,000-token context window and a model size of about 1 GB at 4-bit precision do not establish end-to-end latency, answer quality or operating cost. The material does not specify release dates for each model, licensing terms, or whether all weights and deployment options are available.
Model Cards and Independent Tests
Developers can consult the LFM3 model cards and the agent-harness guide cited by Liquid AI for setup instructions, supported inference engines and implementation details. The supplied source says the models run in Transformers, but does not provide complete release or licensing information.
The next useful evidence would be full benchmark protocols and independent comparisons, along with measurements of memory use, latency and quality after task-specific fine-tuning. Tests on comparable hardware and application tasks would help developers judge whether the reported efficiency gains translate into practical advantages for their workloads.
Key Questions
What is LFM3?
LFM3 is a family of models from Liquid AI built from a shared 1.7-billion-parameter base and adapted for text, vision, audio, multimodal inputs and decision tasks.
Which models are included in the LFM3 family?
The listed models are 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. Liquid AI also lists d1-2B and d1-omni-2B decision models.
Can LFM3 models run on edge devices?
Liquid AI says the family is designed for edge deployment and that developers can fine-tune models on one GPU. The announcement does not independently establish performance, memory use or suitability for particular devices and applications.
Are LFM3’s performance figures independently verified?
Not in the supplied material. The reported speed, parameter-efficiency and benchmark figures come from Liquid AI; independent results and complete evaluation protocols are not provided there.
Where can developers find setup information?
Liquid AI points developers to the LFM3 model cards and an agent-harness guide for setup and inference-engine details. The supplied source says the models are available through Transformers but does not specify all licensing or weight-access terms.
Primary source: Hugging Face · via ThorstenMeyerAI.com
