🔍 Read the full analysis: Explore Specialized Multimodal Model Building 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 built from a shared 1.7-billion-parameter base and adapted for text, vision, audio, multimodal input and decision tasks. The company says the models are designed for single-GPU fine-tuning and edge deployment, but the performance and efficiency figures it cites have not been independently verified 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 of LFM3. The company says developers can fine-tune the models on one GPU and deploy them on devices or other infrastructure, though its performance and efficiency claims have not been independently evaluated in the supplied material.
The 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 additional encoders or decoders for specific modalities where needed. Related multimodal model releases include SenseTime’s open-source multimodal AI model.
Liquid AI says the base 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 despite having 35% fewer parameters, and runs about 25% faster than LFM2-1.7B at an 8,000-token context. It lists a 256,000-token context window for the text model. These are company-reported comparisons; the supplied information does not include independent testing or full evaluation protocols.
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. For the vision system, it reports that a 164-million-parameter 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. The source material does not provide the full vision results or independent confirmation.
Smaller Models for Specialized Deployments
LFM3’s main proposition is that developers can select a model for a specific input type and adapt it, rather than relying on one general-purpose system for every task. If the company’s size, speed and fine-tuning claims hold in practical use, a smaller model could make local processing more feasible for applications that need low latency, operate with limited computing resources or prefer not to send all inputs to a hosted service.
That possibility may be relevant to interactive products, robotics and other settings where systems process audio or video as events unfold. Liquid AI says its audio model can begin processing 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. These are vendor-reported results, not evidence that every application will achieve the same performance. The announcement does not establish real-world cost savings, privacy outcomes or production quality.
A shared base also offers a common starting point for models with different modalities. Developers may be able to build specialized systems while retaining a related underlying architecture. The practical value will depend on the quality of task-specific fine-tuning, deployment hardware, and how the models compare with alternatives on the same tasks.
Products Worth Considering
Sonos Era 100 SL Compact Microphone-Free Speaker, Black
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Sony WH-1000XM6 Wireless Noise-Canceling Over-Ear Headphones, Blue, Bundle
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 LFM2 combined short convolutions with attention to support efficient operation on devices. For LFM3, it reports conducting a hardware-in-the-loop architecture search across 14 device and computing-backend combinations, including phones, a Raspberry Pi 5, laptops, cloud CPUs and GPUs, and Qualcomm’s Hexagon NPU.
The new base changes parts of that design. Liquid AI says its convolution layers use a kernel size of seven, compared with three in LFM2. It also describes head-wise gates that can scale individual attention heads, 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 material does not independently test those design goals.
Liquid AI says the models are available through Transformers and directs developers to individual model cards and an agent-harness guide for supported inference engines and setup details. The information provided does not specify release dates for every model, licensing terms or whether all weights and deployment options are available.
“LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.”
— Liquid AI
Independent Tests and Access Details
The performance figures in the announcement come from Liquid AI. The supplied material does not include independent evaluations, complete benchmark protocols or enough information to compare all results across equivalent hardware and deployment conditions. The vision section is also incomplete in the source, which cuts off before presenting the full comparison.
It remains unclear how much fine-tuning data, time and memory different applications will require, and how model quality and latency change under production workloads. A large context window or a small model file at 4-bit precision does not by itself establish end-to-end speed, output quality or operating cost. The announcement also leaves open licensing, model-by-model release timing and the availability of weights and deployment options.
Model Cards and Comparable Evaluations
Developers can consult the LFM3 model cards and agent-harness guide cited by Liquid AI for implementation instructions, supported inference engines and setup information. The company says the models work through Transformers, but details for particular runtimes and deployment configurations should be checked in the relevant documentation.
The next evidence needed to judge the announcement is access to the models and evaluations on comparable tasks and hardware. Full benchmark methods and results, independent comparisons, and measurements of memory use, latency and quality after task-specific fine-tuning would show whether the reported advantages carry over to practical deployments. Release and licensing details would also clarify which developers can use each model and under what conditions.
Key Questions
What is LFM3?
LFM3 is a family of models introduced by 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 in the LFM3 family?
Liquid AI lists 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. It also lists the decision models d1-2B and d1-omni-2B.
Are the performance claims independently verified?
Not in the supplied material. The comparisons and speed or efficiency figures are reported by Liquid AI; independent evaluations and full benchmark methods are not provided there.
Can developers fine-tune and run LFM3 locally?
Liquid AI says the models are designed for fine-tuning on one GPU and deployment on devices or other infrastructure. Actual hardware needs, runtime support and fine-tuning requirements will depend on the model and application.
When will the models be available, and what are the terms?
The supplied announcement material does not state release dates for each model or specify licensing terms. Liquid AI points developers to model cards and setup documentation for implementation details.
Primary source: Hugging Face · via ThorstenMeyerAI.com
