🔍 Read the full analysis: What Goes Into Building Specialized Models With 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 for text, vision, audio, multimodal input and decision tasks, built from a shared 1.7-billion-parameter base. The company says the models can be fine-tuned on one GPU and deployed on edge devices, but its performance and efficiency 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 input and decision tasks. The company says developers can fine-tune the models on a single GPU and run them on edge devices or other application infrastructure, though the performance and efficiency figures in its original analysis are company-reported, not independently evaluated in the supplied material.
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. Each model branches from LFM3-1.7B-Base, adding modality-specific encoders or decoders where needed.
Liquid AI says the base model uses a 24-layer tri-hybrid architecture that combines short convolutions, sliding-window attention and global attention. It 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. The company lists a 256,000-token context window for the text model and says the base model was trained on 22 trillion tokens, including a final 3.5-trillion-token stage with curated data, more agent traces and tool calls.
The company also reports modality-specific results. Its vision encoder has 164 million parameters; Liquid AI says it 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. For audio, the company says processing can begin while a person is speaking, and that its LiquidCodec tokenizer generates speech at about 87 times real time on one CPU thread on an Apple M4 Max. The supplied material does not include independent evaluations or complete methods for assessing these comparisons.
Small Models for Local Applications
LFM3’s stated aim is to make it easier to adapt a comparatively small model to a defined task and run it near the device or system producing the data. If the reported capabilities hold under practical testing, that could give developers another option for applications that need local processing, low latency or control over data flows, rather than relying entirely on a large hosted model. These are potential benefits, not outcomes established by the announcement.
A shared base also gives developers separate model choices for different input types. Video support and streaming audio could be relevant to interactive assistants, robotics and other systems that interpret changing inputs. Liquid AI says the text model occupies about 1 GB at 4-bit precision and frames the family as suitable for single-GPU fine-tuning. Actual memory needs, tuning effort, operating costs and results will depend on the task, hardware and deployment setup; the source does not provide production measurements or demonstrate privacy gains.
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
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
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 earlier LFM2 models paired short convolutions with attention to support efficient operation on devices. For LFM3, it describes an architecture search involving 14 combinations of devices and computing backends, spanning phones, a Raspberry Pi 5, laptops, cloud CPUs and GPUs, and Qualcomm’s Hexagon NPU.
The company says LFM3 changes the convolution kernel size from three in LFM2 to seven, 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 both the input and output of each block. Liquid AI presents these choices as intended to improve capability per parameter and support longer-context inference; the supplied material does not independently establish those effects.
According to the source, the models are available through Transformers, with model cards and an agent-harness guide offering setup and inference-engine details. Liquid AI characterizes the family as a starting point for specialized systems. Availability of each model, weights and deployment options is not fully specified in the supplied information.
“LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.”
— Liquid AI
Independent Tests Still Needed
The performance, speed and efficiency figures cited here come from Liquid AI. The supplied material does not include independent results, full evaluation protocols or enough detail to compare benchmarks across hardware and deployment conditions. It also begins a discussion of vision benchmarks without providing the full comparison, so that portion cannot be assessed from the information provided.
Other open questions include how much fine-tuning data, time and memory a particular application will require, and how the models perform under production workloads. A stated context window or model size at 4-bit precision does not, on its own, establish end-to-end latency, quality or cost. The source does not specify release dates for every model, licensing terms or the availability of all weights and deployment options.
Model Cards and Comparisons
Developers can consult the LFM3 model cards and the agent-harness guide cited by Liquid AI for implementation instructions and supported inference engines. The supplied source says the models run in Transformers and includes an example for loading the text model, but it does not set out a full release schedule or all access terms.
Further assessment will depend on access to the models and tests that compare them on consistent hardware and application tasks. Useful evidence would include complete benchmark methods and results, independent evaluations, and measurements of memory use, latency and quality after task-specific fine-tuning. Until those are available, the company’s reported gains should be treated as claims rather than independently established performance.
Key Questions
What is LFM3?
LFM3 is a model family from Liquid AI built from a shared 1.7-billion-parameter base and adapted for text, vision, audio, multimodal input and decision tasks.
Which models are included?
The announced family 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 combined modalities; and the decision models d1-2B and d1-omni-2B.
Can developers fine-tune LFM3 on one GPU?
Liquid AI says the models are designed for single-GPU fine-tuning. The supplied material does not give task-specific requirements for training data, time or memory.
Have LFM3’s performance claims been independently verified?
Not in the supplied material. The comparisons and speed figures are reported by Liquid AI, and full evaluation protocols and independent results are not provided.
When will the models be available, and what are the terms?
The source points developers to Transformers, model cards and setup guidance, but does not specify release dates for every model, licensing terms or the availability of all weights and deployment options.
Primary source: Hugging Face · via ThorstenMeyerAI.com
