Liquid AI has released LFM3, a new family of six compact, open-weight AI models designed for text generation, image and video understanding, audio processing, multimodal interaction and specialized decision-making. Announced on October 27, 2026, the lineup shares a common model backbone and targets developers who want to customize and deploy AI on local devices, private infrastructure or cloud systems.
The release centers on a new tri-hybrid architecture that combines three types of sequence-processing layers to balance performance, long-context handling and computational efficiency. Liquid AI says the family was pretrained on 22 trillion tokens and designed to make specialized AI more practical on resource-constrained hardware. The models are available through Hugging Face, the Liquid AI Playground and the company’s API.
Key Takeaways
- Six models: The LFM3 family includes text, vision, audio, omni and two specialized decision models.
- Compact architecture: The lineup spans approximately 1.7 billion to 2.2 billion parameters.
- New design: The tri-hybrid architecture combines gated short convolutions, sliding-window attention and global attention.
- Long context: The LFM3-1.7B text model supports a context window of up to 256,000 tokens.
- Local deployment: The text model can fit in about 1 GB of memory at 4-bit precision, according to Liquid AI.
- Customization: The company says every model can be fine-tuned on a single GPU.
- Availability: Model weights, demonstrations and hosted inference options are available through Liquid AI and its ecosystem partners.
What Is Liquid AI’s LFM3 Model Family?
LFM3 is the latest generation of Liquid Foundation Models from Liquid AI, a company focused on efficient AI systems. Instead of releasing one large, general-purpose model for every task, Liquid AI has built a family of specialized models around a shared foundation.
The idea is to let developers select the model that best matches their application. A text-based agent does not necessarily need a full audio-generation pipeline, while a voice assistant needs more than text generation alone.
The six released models are:
| Model | Primary purpose | Main capabilities |
|---|---|---|
| LFM3-1.7B | Text and agent workflows | Text generation, instruction following and tool-oriented tasks |
| LFM3-VL-2B | Vision | Text, image and video understanding |
| LFM3-Audio-2B | Speech and audio | Audio understanding, text and speech generation |
| LFM3-Omni-2B | Multimodal interaction | Text, images, video and audio, with text or speech output |
| d1-2B | Vision-based decisions | Specialized, calibrated decisions from text and visual inputs |
| d1-Omni-2B | Multimodal decisions | Decision-making across text, visual and audio inputs |
The decision models are specialized versions of the vision and omni models. They are designed to return calibrated answers rather than operate as conventional conversational assistants.
This modular approach could help developers build applications with smaller components instead of relying on a single large model for every operation.
How the Tri-Hybrid Architecture Works
The central technical change in LFM3 is its tri-hybrid architecture. The LFM3-1.7B-Base model contains 24 layers that combine three different mechanisms.
| Component | Number of layers | Purpose |
|---|---|---|
| Gated short convolutions | 13 | Efficiently process local sequence patterns |
| Sliding-window attention | 5 | Focus attention on a limited local context |
| Global attention | 6 | Connect information across longer sequences |
Short convolutions help process nearby information efficiently. Sliding-window attention lets the model focus on a defined portion of the sequence, while global attention allows it to relate information across a much longer context.
Combining these mechanisms is intended to balance computational cost and model capability. A model using global attention everywhere can become expensive as its context grows, while a system relying only on local processing may struggle to connect information that appears far apart.
Liquid AI says its architecture search evaluated 14 device and backend combinations, expanding beyond the limited hardware configurations used for its previous generation. These included smartphones, laptops, a Raspberry Pi 5, cloud CPUs, GPUs and Qualcomm’s Hexagon neural processing unit.
The company also introduced changes to attention gating, positional encoding and normalization. These adjustments aim to improve model quality and quantization behavior without imposing a significant inference-speed penalty.
Why Local AI Deployment Matters
Many AI applications currently depend on cloud-hosted models. While this can provide access to powerful computing resources, it also introduces network latency, recurring inference costs and potential concerns about sending sensitive information to external services.
Smaller models can offer another option. LFM3 is designed to run on local devices, allowing developers to deploy AI closer to where information is created or processed.
Liquid AI says the LFM3-1.7B text model fits in approximately 1 GB of memory at 4-bit precision. This makes it a potential candidate for smartphones, laptops and embedded systems, although actual memory requirements depend on the runtime, context length and workload.
The text model also supports a 256,000-token context window. That can be useful for applications involving lengthy documents, extended conversations or agent workflows that need to preserve substantial context.
However, fitting model weights into a limited amount of memory is not the same as fitting the entire application into that amount. Runtime overhead, the key-value cache, other components and longer context windows require additional resources.
Vision, Audio and Omni Models Expand the Use Cases
LFM3 extends beyond text generation through specialized vision and audio models.
LFM3-VL-2B accepts text, images and video. Its video capability allows applications to examine sampled frames and answer questions about visual content. This could support document analysis, visual inspection and video-understanding tools.
LFM3-Audio-2B is designed for speech and audio interactions, including audio input and speech generation. Liquid AI’s architecture includes a dedicated audio encoder and LiquidCodec, its audio-tokenization technology, to support low-latency speech processing.
LFM3-Omni-2B combines text, image, video and audio understanding in a single model. It can respond using text or speech, enabling applications such as voice assistants that can discuss a photograph or answer questions about a video and its soundtrack.
The advantage of an omni model is that it can process different types of information within one model pipeline. Developers may not need to connect separate speech-recognition, vision, language and speech-generation systems for every application.
That does not mean the models will outperform larger systems in every task. Performance will depend on the workload, hardware, prompting approach and evaluation method.
What the Decision Models Add
The two d1 models are designed for applications where the main objective is to make a structured decision rather than generate a long answer.
For example, a business application could use a vision-based decision model to categorize an image, while an agent could use a specialized model to evaluate an input against a predefined set of outcomes.
This approach can be useful in high-volume workflows that need many repeated evaluations. A smaller decision model may reduce the computational overhead of calling a large general-purpose model for every classification or scoring task.
Liquid AI has reported improvements on selected decision benchmarks for the new models. As with other vendor-reported results, those numbers should be evaluated against independent testing and the specific requirements of a real application.
Performance, Fine-Tuning and Availability
Liquid AI says LFM3-1.7B outperforms its previous LFM2.5-2.6B model on its evaluation suite while using 35% fewer parameters. The company also reports that LFM3-1.7B is approximately 25% faster than LFM2-1.2B on 8,000-token contexts under its tested conditions.
These are company-reported measurements, and performance varies by benchmark and hardware. They should not be interpreted as proof that LFM3 is faster or more capable than every competing model.
Developers can access the model weights through Hugging Face, experiment through the Liquid AI Playground and use hosted inference through the Liquid AI platform.
The family also supports deployment through tools and runtimes in the wider ecosystem, including llama.cpp, MLX, vLLM, SGLang, Ollama and Transformers, with support depending on the individual model and configuration.
Liquid AI says the models can be fine-tuned on a single GPU, allowing businesses to adapt them to specialized terminology, proprietary datasets and task-specific requirements.
Developers should check each model’s individual license before commercial use. Open-weight availability does not mean every model has identical licensing conditions.
The Bigger Picture
LFM3 reflects a growing effort to make AI more practical beyond large cloud-based systems. Instead of measuring progress only through parameter counts, model developers are increasingly focusing on how much useful capability can run within a particular memory budget, power envelope and latency target.
The shared-backbone approach also gives developers a choice between specialized models and broader multimodal functionality. Businesses could use local models for privacy-sensitive tasks, combine them with cloud systems for complex requests or fine-tune smaller models for narrow operational needs.
The real test will be how LFM3 performs in independent evaluations and production deployments. If its efficiency claims translate into reliable applications, the family could make local AI more accessible to developers building assistants, automation tools and multimodal software.
FAQs
1. What is Liquid AI LFM3?
LFM3 is a family of six open-weight AI models covering text, vision, audio, multimodal interaction and specialized decision-making. The models share a common backbone and target efficient deployment across devices and cloud infrastructure.
2. What is the tri-hybrid architecture?
It combines gated short convolutions, sliding-window attention and global attention in one model. The design aims to balance computational efficiency with the ability to process local patterns and longer-range relationships.
3. Can LFM3 run on a smartphone?
Liquid AI designed the family for local deployment, and the text model can fit in approximately 1 GB of memory at 4-bit precision. Actual device compatibility and performance depend on hardware, context length and the selected runtime.
4. What is LFM3-Omni-2B?
LFM3-Omni-2B is a multimodal model that accepts text, images, video and audio, and can respond with text or speech. It is intended for applications that need multiple input modalities within one model.
5. Is LFM3 open source?
Liquid AI describes LFM3 as an open-weight model family, meaning developers can access model weights and deploy them under the applicable terms. However, licensing conditions differ across the family, so developers should check the specific model license before commercial deployment.
Looking Ahead
The next milestone for LFM3 will be adoption by developers and independent testing across mobile devices, laptops, embedded systems and cloud infrastructure. Performance in real applications—not just benchmark averages—will determine whether the models can deliver meaningful advantages in latency, memory consumption and operating cost.
As compact multimodal models improve, businesses may increasingly combine local AI with cloud-based services rather than choosing one approach exclusively. Liquid AI’s LFM3 family is a step toward that model, offering specialized capabilities on a shared architecture designed for efficient deployment and customization.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
