🔍 Read the full analysis: Build AI That Handles Multiple Modalities 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 model family 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 supplied announcement provides no independent evaluations of its performance claims.
Liquid AI has introduced LFM3, a family of models based on a shared 1.7-billion-parameter foundation and adapted for text, images, video, audio and decision tasks. The company says developers can fine-tune the models on one GPU and deploy them on edge devices or other infrastructure, positioning the release for teams building specialized AI systems; the performance figures provided so far are company-reported, not independently evaluated, as discussed in the original analysis.
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 the LFM3-1.7B-Base, adding modality-specific components where required.
Liquid AI describes that base 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 with curated data containing more agent traces and tool calls. The text model is listed with a 256,000-token context window; Liquid AI says it occupies 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, and that it runs about 25% faster than LFM2-1.7B at an 8,000-token context. Other reported figures include a 164-million-parameter vision encoder that Liquid AI says 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. The source material does not provide independent test results or full evaluation protocols for these claims.
LFM3’s design gives developers a set of modality-specific options rather than requiring one model to handle every input type. If the company’s reported size and speed figures hold in real applications, smaller organizations could have more choices for running specialized models close to where data is produced, including on devices or local infrastructure. That may be relevant to applications where latency, connectivity or data handling affect system design.
The announced capabilities suggest possible uses in interactive assistants, robotics and systems that process video or streaming speech. Liquid AI 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. Those are vendor claims, however; the announcement does not establish production performance, cost savings, privacy outcomes or suitability for any particular deployment.
The single-GPU fine-tuning proposition may lower the hardware barrier for teams adapting a model to a narrower task. But the parameter count and context window alone do not show how much memory, data, tuning time or engineering work a given application will need. Real-world quality and operating costs remain questions for testing.
Popular right now
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
KitchenAid Artisan Series 5-Quart Tilt-Head Stand Mixer, Dried Rose (KSM150PS)
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 paired short convolutions with attention to support efficient operation on devices. For LFM3, it reports that a hardware-in-the-loop architecture search 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 new base changes the convolution kernel size from three in LFM2 to seven, according to Liquid AI. The company 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. It says these choices are intended to improve capability per parameter and support longer-context inference; the supplied material does not independently establish those effects.
Liquid AI says the models are available through Transformers and directs developers to model cards and an agent-harness guide for setup, inference engines and implementation information. The source material does not specify licensing terms or confirm the availability of all model weights and deployment options.
““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 performance, efficiency and accuracy figures in the supplied material come from Liquid AI. It does not include independent evaluations, complete benchmark methods or enough detail to compare results across different hardware and deployment conditions. A portion of the source’s discussion of vision results is also incomplete, so the full benchmark picture is not available here.
It is not yet clear how much fine-tuning data, time or memory a typical application will require, or how the models perform under production workloads. A stated 256,000-token context window and a model size at 4-bit precision do not, by themselves, establish response quality, end-to-end latency or operating cost. The announcement also leaves licensing, release timing for individual models, and the availability of all weights and deployment options unspecified.
Model Cards and Deployment Evidence
Developers can consult the LFM3 model cards and the agent-harness guide cited by Liquid AI for setup instructions, supported inference engines and other implementation details. The company says the models run in Transformers, but the supplied material does not provide a release timetable for every model or resolve licensing and availability questions.
The next useful evidence will be full benchmark methods and results, independent comparisons on comparable hardware, and measurements of memory use, latency and quality after task-specific fine-tuning. Those results would help developers judge whether the company’s reported capabilities translate into practical gains for their own applications.
Key Questions
What is LFM3?
LFM3 is Liquid AI’s model family built from a shared 1.7-billion-parameter base, with versions adapted for text, vision, audio, multimodal input and decision tasks.
Which input types do the LFM3 models support?
Liquid AI lists models for text, text with images and video, text with audio, and combined text, image, video and audio input. The family also includes two decision models.
Can LFM3 run on edge devices?
Liquid AI says the models are intended for deployment on edge devices as well as other infrastructure, and that they can be fine-tuned on one GPU. Actual hardware requirements and performance will depend on the model and application.
Are LFM3’s performance claims independently verified?
Not in the supplied material. The speed, accuracy and efficiency figures are reported by Liquid AI, which does not provide independent evaluations or complete benchmark protocols there.
Where can developers find setup information?
Liquid AI points developers to the LFM3 model cards and an agent-harness guide for setup and supported inference engines. The announcement says the models are available through Transformers, but does not state all release and licensing details.
Primary source: Hugging Face · via ThorstenMeyerAI.com
