🔍 Read the full analysis: Build 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 derived from a shared 1.7-billion-parameter base and adapted for text, vision, audio, multimodal inputs and decision tasks. The company says the models can be fine-tuned on one GPU and deployed on edge devices, but the performance and efficiency figures it cites have not been independently evaluated in the supplied material.
Liquid AI has introduced LFM3, a family of models derived 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 application infrastructure, though its performance and efficiency figures are company-reported and have not been independently assessed in the supplied material.
The announced 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 text, images, video and audio. Liquid AI also lists two multimodal decision models, d1-2B and d1-omni-2B, which the company describes as producing calibrated answers. Each model branches from LFM3-1.7B-Base, with modality-specific encoders or decoders added where needed.
Liquid AI says the base model uses a 24-layer tri-hybrid architecture combining 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. It also says the model is about 25% faster than LFM2-1.7B at an 8,000-token context, and that the text model supports a 256,000-token context window.
The company says 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. Liquid AI says the text model occupies about 1 GB at 4-bit precision and is suited to fine-tuning on one GPU. The announcement points developers to model cards for inference-engine support and setup instructions, and says the models are available through Transformers.
LFM3’s design gives developers several task-specific model options while keeping a shared base architecture. That could be useful for teams that want a model matched to their input type rather than a single general-purpose system, particularly where applications handle audio or video as they arrive. Liquid AI says its audio system can start processing while a person is speaking, a feature that may matter for interactive applications.
The focus on fine-tuning and deployment on local devices is relevant to products with latency, connectivity or data-handling constraints. If the company’s size and speed figures hold in practical settings, some applications may be able to run without depending on a large hosted model for every task. But the announcement does not establish actual reductions in cost, latency or data exposure. Those outcomes depend on the hardware, workload and implementation.
The lineup’s breadth also makes evaluation important: a model that handles several input types is not automatically more accurate or efficient for a particular job. Developers will need to test each option against their own tasks, compare it with alternatives under similar conditions, and account for the extra encoders or decoders used for different modalities.
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
Sonos Era 100 SL Compact Microphone-Free Speaker, Black
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
As an affiliate, we earn on qualifying purchases.
From LFM2 to LFM3’s Architecture
LFM3 follows Liquid AI’s LFM2 family. The company says its earlier work paired short convolutions with attention to support efficient operation on devices. For LFM3, it describes a hardware-in-the-loop architecture search covering 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 new base model changes several parts of that design. Liquid AI says its convolution layers use a kernel size of seven, compared with three in LFM2, and that head-wise gates can scale individual attention heads. The company also describes partial rotary position embeddings in global attention layers and normalization at both the input and output of each block. These are design details and stated intentions; the supplied material does not independently establish how each change affects performance across deployments.
Liquid AI also reports modality-specific results. It says its 164-million-parameter vision encoder matched 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. For speech generation, it says LiquidCodec runs at about 87 times real time on one CPU thread on an Apple M4 Max. The source does not provide independent evaluations or complete benchmark protocols for these claims.
“LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.”
— Liquid AI
Independent Benchmarks Still Needed
The comparisons and efficiency figures in the supplied material come from Liquid AI. It does not include independent test results or enough detail about evaluation methods, hardware and deployment conditions to establish how the models compare across real applications. The source also begins describing vision benchmark results but is cut off before supplying the full comparison.
It remains unclear how much training data, time and memory different fine-tuning tasks will require, or how the models perform under production workloads. A 256,000-token context window and a reported 1 GB size at 4-bit precision do not, by themselves, show end-to-end latency, output quality or operating cost. The source does not specify the release date for each model, licensing terms, or whether all weights and deployment options are available.
The decision models’ description as producing “calibrated answers” is also a company characterization; the supplied information does not include a calibration metric or independent result. Likewise, possible uses in robotics, assistants and other interactive systems are suggested by the models’ capabilities, not confirmed evidence of deployed performance.
Model Cards and Evaluations
Developers can consult the LFM3 model cards and Liquid AI’s agent-harness guide for information about setup, supported inference engines and implementation. The supplied material says the models work through Transformers, but further deployment details will depend on the individual model documentation.
The next useful evidence would include complete benchmark protocols and results, independent comparisons on equivalent hardware, and measurements of memory use, latency and quality after fine-tuning. Details on licensing, weight availability and the release status of each model would also clarify who can test and deploy the family. Until those details are available, the announcement establishes the lineup and its intended design, not its performance in any particular production setting.
Key Questions
What is LFM3?
LFM3 is a family of Liquid AI models 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 lineup includes LFM3-1.7B for text; LFM3-VL-2B for text, images and video; LFM3-Audio-2B for text and audio; LFM3-Omni-2B for text, images, video and audio; and the d1-2B and d1-omni-2B decision models.
Can LFM3 models be fine-tuned on one GPU?
Liquid AI says the models are designed for single-GPU fine-tuning. The supplied material does not specify how much memory, training time or data a particular fine-tuning task will need.
Have LFM3’s performance claims been independently verified?
The supplied source provides company-reported figures but no independent evaluations or complete benchmark protocols. Results in specific applications remain to be established.
When will the models be available, and what are the licensing terms?
The source says the models are available through Transformers and directs developers to model cards, but it does not give individual release dates or licensing terms, or clarify the availability of all weights and deployment options.
Primary source: Hugging Face · via ThorstenMeyerAI.com
