🔍 Read the full analysis: Customize Multimodal AI For Specific Tasks Using LFM3 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 adapted from a shared 1.7-billion-parameter base for text, vision, audio, multimodal inputs and decision tasks. The company says the models are designed for single-GPU fine-tuning and edge deployment, 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, images, video, audio and decision tasks, as detailed in the original analysis. The company says developers can fine-tune the models on a single GPU and deploy them on edge devices or other infrastructure, though the performance and efficiency figures it cites have not been independently evaluated in the supplied material.
The announced 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 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. Each model branches from LFM3-1.7B-Base, with modality-specific components added where needed.
Liquid AI describes the base as a 24-layer tri-hybrid architecture that combines short convolutions, sliding-window attention and global attention. The company says the 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 text model is listed with a 256,000-token context window; Liquid AI also says it fits in about 1 GB at 4-bit precision.
The company reports several comparisons with earlier models. It says LFM3-1.7B outperforms its LFM2.5-2.6B predecessor while using 35% fewer parameters, and runs about 25% faster than LFM2-1.7B at an 8,000-token context. Liquid AI 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. These are company-reported results; the source material does not provide independent evaluations or full test protocols.
Smaller Models for Local Tasks
LFM3’s pitch is that developers may be able to adapt a relatively small model to a specific task and run it close to where data is produced. If its size and speed claims hold in practical use, that approach could reduce dependence on larger hosted models for applications that need local processing, low latency or control over data flows. The announcement does not establish that those benefits have been achieved in production.
A family of models organized by input type could also give teams a choice between text, vision, audio and combined systems, rather than relying on a single model for every task. Video input and streaming audio may be relevant to interactive applications such as assistants and robotics. Those are potential uses suggested by the model design, not demonstrated outcomes. Real-world quality, deployment costs and privacy effects remain unconfirmed by the information provided.
Products Worth Considering
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From LFM2 to LFM3
Liquid AI presents LFM3 as the next generation of its Liquid Foundation Models, following the LFM2 family. The company says its earlier work paired short convolutions with attention to support efficient operation on devices. For LFM3, it says an architecture search using hardware in the loop covered 14 combinations of devices and computing backends, from phones and a Raspberry Pi 5 to laptops, cloud CPUs and GPUs, and Qualcomm’s Hexagon NPU.
The company says the new base uses convolution kernels of size seven, compared with three in LFM2, and adds head-wise gates that can scale individual attention heads. It also describes partial rotary position embeddings in global attention layers and normalization at the input and output of each block. Liquid AI says these choices are intended to improve capability per parameter and support longer-context inference; the supplied material does not include independent tests of those effects.
Liquid AI says the models are available through Transformers and directs developers to model cards for inference-engine and setup details. The announcement frames the shared base as a starting point for task-specific adaptation, rather than claiming that one model configuration suits every deployment.
“LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.”
— Liquid AI
Benchmarks and Access Still Need Detail
The supplied material contains company-reported performance and efficiency figures, but no independent results, full evaluation protocols or sufficient detail to compare the benchmarks across hardware and deployment conditions. It also starts to describe vision-model benchmark results but does not include the full comparison. The figures therefore do not establish how the models will perform on a particular application or device.
It is not clear how much fine-tuning data, time or memory different tasks will require, or how the models perform under production workloads. A long context window and a reported model size at 4-bit precision do not alone establish end-to-end latency, answer quality or operating cost. The supplied announcement also does not specify licensing terms, release timing for each model, or whether all weights and deployment options are available.
Model Cards and Independent Tests
Developers can consult the LFM3 model cards and agent-harness guide cited by Liquid AI for setup instructions and supported inference engines. The supplied source says the models run in Transformers and points to an example for loading the text model. Those materials may clarify implementation requirements, but the information provided here does not settle availability, licensing or deployment conditions for every model.
Further assessment will depend on access to the models and testing on comparable hardware and application tasks. Useful next evidence would include full benchmark methods and results, independent comparisons, and measurements of memory use, latency and quality after fine-tuning. Until those details are available, developers cannot determine from the announcement alone whether LFM3 will meet the needs or costs of a specific deployment.
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 did Liquid AI announce?
The named 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. The company also lists d1-2B and d1-omni-2B decision models.
Can LFM3 be fine-tuned on one GPU?
Liquid AI says the models are designed for fine-tuning on a single GPU. The supplied material does not specify the data, memory or compute requirements for particular fine-tuning jobs.
Are the reported LFM3 benchmarks independently verified?
Not in the supplied material. The performance and efficiency figures are reported by Liquid AI; independent evaluations and complete testing methods are not provided there.
When are the models available, and what are the license terms?
The source says the models are available through Transformers and points developers to model cards, but it does not give release dates for each model or licensing terms. Those details should be checked in Liquid AI’s model documentation.
Primary source: Hugging Face · via ThorstenMeyerAI.com
