TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
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 use and decision tasks. The company says the models are designed for fine-tuning on one GPU and deployment on edge devices; its performance and efficiency figures are company-reported.
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. The company says developers can fine-tune each model on a single GPU and deploy it on the device or infrastructure their application uses, an approach aimed at bringing specialized multimodal systems to smaller deployments.
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. Two decision models, d1-2B and d1-omni-2B, produce what Liquid AI describes as calibrated answers. Each branches from LFM3-1.7B-Base, with modality-specific encoders or decoders where needed.
Liquid AI says its new base model uses a 24-layer tri-hybrid design 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 context window of 256,000 tokens.
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. For vision, Liquid AI reports that 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. It says its audio system 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 company-reported figures; the supplied material does not include independent evaluations.
Liquid AI · Model family overview
Build Specialized Multimodal Models With LFM3
One shared base. Specialized paths for text, vision, audio and decision tasks. Liquid AI says the LFM3 family is designed for single-GPU fine-tuning and deployment on edge devices or application infrastructure. Reported performance figures come from the company.
One base, six specialized models
Each model branches from LFM3-1.7B-Base, adding modality-specific encoders or decoders where needed.
LFM3-1.7B
Text model for language tasks, with a reported 256,000-token context window.
Text in · Text outLFM3-VL-2B
Handles text, images and video for visual understanding tasks.
Text · Image · VideoLFM3-Audio-2B
Combines text and audio, with support for streaming audio input.
Text · AudioLFM3-Omni-2B
Brings text, image, video and audio inputs together in one model.
Text · Image · Video · Audiod1-2B
Decision model that Liquid AI describes as producing calibrated answers.
Decision tasksd1-omni-2B
Decision model for multimodal inputs; calibration claims need evaluation.
Multimodal decisionsA tri-hybrid design for capability per parameter
Liquid AI says hardware-in-the-loop searches informed the architecture across 14 device and backend combinations.
The 24-layer base combines short convolutions, sliding-window attention and global attention. The design builds on LFM2’s convolution-and-attention approach, with changes intended to support longer context and efficient inference.
Other described choices include kernel size seven convolutions, head-wise gates, partial rotary position embeddings in global attention, and normalization at each block’s input and output.
Why a smaller specialized model?
The proposed benefit is local adaptation and inference for applications with specific inputs, latency needs or data flows.
Designed for smaller deployments
The company says models can be fine-tuned on a single GPU and deployed where an application runs. It also says the text model fits in about 1 GB at 4-bit precision.
- Streaming audio can begin processing while a person speaks.
- The vision encoder is reported at 164M parameters.
- Model access is described through Transformers and individual model cards.
Real-world results need testing
Smaller size and company benchmarks do not establish quality, cost or reliability in a production application.
- Independent results and complete benchmark methods
- Memory, latency and quality on comparable hardware
- Fine-tuning data, time and memory requirements
A practical evaluation path
For edge use, the deciding evidence is how a tuned model performs on the target task and hardware.
Follow the evidence chain
Architecture choices are testable; practical benefits depend on independent, task-specific results.
What developers should know
What is LFM3?
Liquid AI’s model family, built from a shared base and adapted for text, vision, audio, multimodal and decision tasks.
Which inputs can it handle?
It depends on the model: LFM3-VL-2B supports text, images and video; LFM3-Omni-2B combines those with audio.
Can teams fine-tune it locally?
Liquid AI says each model can be fine-tuned on a single GPU. Requirements vary with data, configuration and task.
Are the performance claims independent?
No independent evaluations are included in the supplied material. Treat performance figures as company-reported until reproduced.
Specialized Models for Edge Devices
LFM3’s central proposition is that teams may be able to adapt a relatively small model to a particular task and run it close to where data is produced. If the company’s size and speed claims hold in practical deployments, that could reduce reliance on large hosted models for applications that need local processing, low latency or tighter control over data flows. Liquid AI says its text model fits in about 1 GB at 4-bit precision, and frames the family as suitable for single-GPU fine-tuning.
A shared base also gives developers a choice of model by input type rather than requiring one general-purpose system to handle every task. Video input in the vision model and streaming audio support could matter for robotics, assistants and other interactive applications. Those uses remain possibilities suggested by the design; the announcement does not establish real-world performance, cost savings or privacy outcomes.
Products Worth Considering
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
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 work paired short convolutions with attention to run efficiently on devices. For LFM3, it says a hardware-in-the-loop architecture search covered 14 combinations of devices and computing backends, ranging from phones and a Raspberry Pi 5 to laptops, cloud CPUs and GPUs, and Qualcomm’s Hexagon NPU.
The LFM3 base changes that design in several ways. Its convolution layers use a kernel size of seven, up from three in LFM2, while 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. Liquid AI says these choices are intended to support capability per parameter and longer-context inference. The models are available through Transformers, according to the supplied material, which points developers to individual model cards for supported inference engines and setup details.
“LFM3 is our new generation of Liquid Foundation Models (LFMs), built to be the starting point for specialized intelligence.”
— Liquid AI
Independent Results Remain Unreported
The performance comparisons and efficiency figures in the supplied material come from Liquid AI. It does not provide independent test results, full evaluation protocols or enough detail here to compare the reported benchmarks across different hardware and deployment conditions. The source also begins to describe vision-model benchmark results but is cut off before giving the full comparison.
It is also unclear how much fine-tuning data, time and hardware memory different applications will require, or how the models perform under production workloads. A 256,000-token context window and a reported model size at 4-bit precision do not by themselves establish end-to-end latency, quality or operating cost. The announcement does not specify release dates for each model, licensing terms, or the availability of all weights and deployment options.
Model Cards and Deployment Details
Developers can consult the LFM3 model cards and 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 shows an example for loading the text model. Further scrutiny will depend on access to the models and evaluations that test them 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 task-specific fine-tuning. Those details would help determine whether the shared-backbone design delivers the practical edge-device benefits the company describes.
Where I land
My assessment is that LFM3 is a concrete attempt to make specialized multimodal models easier to adapt and run on smaller systems. A common base, distinct modality models and single-GPU fine-tuning are relevant design choices for teams that cannot rely on a large hosted model for every task. The announcement is specific about architecture and intended deployment, which makes its claims testable.
The strongest counterargument is that model size and company benchmarks do not show whether the models will be accurate, fast or economical in a real application. I would revise my assessment upward if independent evaluations reproduced the reported results across representative edge hardware and showed that fine-tuned versions remain reliable on actual user tasks. Results that fell short under those conditions would weaken the case for choosing LFM3 on the basis of this announcement.
Key Questions
What is LFM3?
LFM3 is Liquid AI’s new model family, built from a shared base and adapted for text, vision, audio, multimodal and decision tasks.
Which inputs can the models handle?
The family includes models for text, images, video and audio. The exact inputs and outputs depend on the model: for example, LFM3-VL-2B accepts text, images and video, while LFM3-Omni-2B combines those inputs with audio.
Can developers fine-tune LFM3 locally?
Liquid AI says each model can be fine-tuned on a single GPU and describes the family as suitable for deployment on the hardware where an application runs. Requirements for a particular fine-tuning job will depend on its data and configuration.
Are LFM3’s performance claims independently verified?
The supplied announcement attributes the performance and efficiency figures to Liquid AI. It does not provide independent verification or full benchmark details.
Source: Hugging Face
