🔍 Read the full analysis: Build Multimodal AI Models Tailored To Your Needs 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 text, vision, audio, multimodal and decision models built from a shared 1.7-billion-parameter base. 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 evaluated in the supplied material.
Liquid AI has introduced LFM3, a family of models for text, images, video, audio and decision tasks, built from a shared 1.7-billion-parameter base. The company says developers can adapt the models with a single GPU and deploy them on edge devices or other infrastructure, a pitch aimed at teams building specialized AI applications, as described in the original analysis of LFM3; the cited performance and efficiency results are company-reported, not independently verified in the material provided.
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 and uses modality-specific components where needed.
Liquid AI describes the base model as a 24-layer tri-hybrid architecture combining 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. It reports a 256,000-token context window for the text model and says that model fits in about 1 GB at 4-bit precision.
The company reports that LFM3-1.7B outperforms its LFM2.5-2.6B predecessor while using 35% fewer parameters. It also says LFM3-1.7B runs about 25% faster than LFM2-1.7B at an 8,000-token context. For vision, Liquid AI says its 164-million-parameter 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. Those figures depend on the company’s tests; full protocols and independent results were not included in the supplied source.
A Smaller Route to Specialized AI
LFM3’s central proposition is that developers could start with a shared, comparatively small base, then choose a model suited to their input type and adapt it to a particular task. If the reported size and speed figures hold in real deployments, that approach could help some teams run AI closer to where data is produced instead of relying entirely on large hosted systems.
That could be relevant for applications where latency, local processing or infrastructure constraints shape the design, including interactive assistants and devices that process audio or video. Liquid AI says its audio system can begin processing while someone 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 similar performance or lower costs.
The distinction between a model family and a proven deployment benefit matters. A short context run, a particular accelerator or a controlled benchmark may not reflect production workloads. The announcement does not establish that LFM3 will improve privacy, reduce operating expenses or match larger systems on a given task. Those outcomes depend on the application, data, hardware and evaluation method.
Products Worth Considering
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
Sony WH-1000XM6 Wireless Noise-Canceling Over-Ear Headphones, Blue, Bundle
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Portable Waterproof Speaker, Black
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 combined short convolutions with attention to support efficient device execution. For LFM3, it describes an architecture search that 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 company says LFM3 changes several design elements. Its convolution layers use a kernel size of seven, compared with three in LFM2; it also describes gates that scale individual attention heads, partial rotary position embeddings in global attention layers, and normalization at both the input and output of each block. Liquid AI says those choices aim to improve capability per parameter and support longer-context inference. These are explanations of design intent, not independent findings about results.
The supplied material says the models are available through Transformers and directs developers to individual model cards for supported inference engines and setup information. Liquid AI describes the family as a starting point for specialized systems, rather than a single model intended to handle every task in the same way.
“LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.”
— Liquid AI
Independent Testing Is Still Needed
The supplied material does not include independent evaluations, full benchmark protocols or complete results for the cited comparisons. A vision benchmark discussion in the source is incomplete, and the hardware and conditions behind several performance claims are not detailed enough here to compare them across deployments.
It is also unclear how much fine-tuning data, time and memory different applications will require, or how the models perform under production workloads. A 256,000-token context window and a reported size at 4-bit precision do not, by themselves, establish model quality, end-to-end latency or operating cost. The announcement material also does not specify release dates for each model, licensing terms, or whether all weights and deployment options are available.
Those gaps mean the claims should be treated as Liquid AI’s reported results until tested against comparable models on the same hardware and tasks. Practical outcomes may vary with inference software, quantization, input type and fine-tuning choices.
Model Cards and Deployment Evidence
Developers can consult the LFM3 model cards and the agent-harness guide cited by Liquid AI for implementation details, supported inference engines and setup instructions. The supplied source says the models run in Transformers and includes an example for loading the text model, but it does not give a complete schedule for availability or clarify licensing.
The next useful evidence would be full benchmark methods and results, independent comparisons on equivalent hardware, and measurements of memory use, latency and quality after task-specific fine-tuning. Those tests would help determine whether the reported advantages carry over from company evaluations to particular applications. Until then, developers considering LFM3 will need to check the model-specific documentation and evaluate the relevant version on their own workloads.
Key Questions
What is LFM3?
LFM3 is a family of AI models introduced by Liquid AI. It includes models for text, vision, audio, combined modalities and decision tasks, based on a shared 1.7-billion-parameter base.
Which inputs do the LFM3 models support?
The listed models cover text, images, video and audio in different combinations. LFM3-1.7B is for text; the VL model handles text, images and video; the Audio model handles text and audio; and Omni combines all four input types.
Has LFM3’s performance been independently verified?
Not in the supplied material. The speed, accuracy and efficiency figures cited are from Liquid AI, and the source does not provide independent evaluations or full benchmark protocols.
Can developers fine-tune and run LFM3 locally?
Liquid AI says the models are designed for fine-tuning on a single GPU and deployment on edge devices or other infrastructure. The exact hardware, memory and setup needs will depend on the model and application; developers should check the model cards for available implementation details.
Primary source: Hugging Face · via ThorstenMeyerAI.com
