A family of models, not one recipe for every modality
Liquid AI introduced LFM3 as a new generation of its Liquid Foundation Models family. The release combines a shared base of about 1.7 billion parameters with derived models covering text, images, video, audio, and decision tasks. It is not a single model that handles all of these in the same way: available modalities depend on the variant you choose.
The announcement describes six models: LFM3-1.7B for text; LFM3-VL-2B for text, images, and video; LFM3-Audio-2B for text and audio; LFM3-Omni-2B for text, image, video, and audio input; and two decision models, d1-2B and d1-omni-2B. The d1 models produce structured answers or scores for defined questions rather than focusing on free-form text generation.
The practical distinction matters: “multimodal” does not mean that every variant accepts every modality. For example, the vision model processes video, while audio input is handled by the audio and Omni variants. The family shares an architectural starting point, but each model’s specific capabilities depend on the encoders, decoders, and tasks it includes.
Inputs and outputs listed for each variant
This summary reflects the modalities described by Liquid AI; it is not an independent evaluation of performance.
| Model | Listed inputs | Listed outputs or function |
|---|---|---|
| LFM3-1.7B | Text | Text |
| LFM3-VL-2B | Text, images, and video | Text |
| LFM3-Audio-2B | Text and audio | Text, speech, or both |
| LFM3-Omni-2B | Text, images, video, and audio | Text and speech |
| d1-2B | Text, images, and video | Calibrated answers to structured questions |
| d1-omni-2B | Text, images, video, and audio | Calibrated answers to structured questions |
A shared architecture and changes to the text model
The LFM3-1.7B-Base model combines three types of layers: gated short convolutions, sliding-window attention, and global attention. Liquid AI says the design has 24 layers: 13 convolutional layers, five sliding-window attention layers with a 256-token window, and six global attention layers. The company presents this combination as a way to balance inference cost with capability over long contexts.
The announcement also describes head-wise gated attention, partial RoPE in global attention layers, and normalization at both the input and output of each block. These are architectural choices; by themselves, they do not prove that a model is more accurate, reliable, or safe. Liquid AI’s stated motivation is to improve efficiency and model behavior, including effects that can influence quantization.
According to Liquid AI, the base model was pretrained on 22 trillion tokens and then received an additional 3.5-trillion-token stage using curated data, with a larger share of tool-use examples and agent traces. The announcement lists a 256,000-token context window for the text model. That figure describes the advertised limit; it does not guarantee that every task will retain the same quality at that length or that the hardware can process such a context at a practical cost.
What “specialization” means and what can be fine-tuned
In this announcement, specialization means adapting a model to an application’s data, vocabulary, tasks, or tools through fine-tuning. Liquid AI provides examples using different methods: SFT for extracting structured data from documents, DPO for answering from retrieved documents or abstaining, GRPO for choosing tools, and CPT for learning from text in another language. It also includes exercises for vision, audio, and Omni models.
The announcement says each model can be fine-tuned on a single GPU and presents notebooks with workflows for data preparation and evaluation. This is a company claim about its training approach, not enough information to estimate the cost of a particular project. Memory requirements, fine-tuning time, and performance will depend on the method, numeric precision, dataset size and format, context length, and hardware setup.
One example adapts LFM3-VL-2B for robotics, with models for orchestration, navigation, and manipulation. Liquid AI reports success rates on evaluation suites and generation times on specific hardware. Those results concern the configurations and tests presented by the company; they do not establish that a custom fine-tune will achieve similar results or transfer unchanged to another robot or environment.
A cautious workflow for evaluating a specialization
The announcement provides fine-tuning examples, but every application needs separate checks to establish whether a model works with its data and operating conditions.
- 01Define a narrow task and an observable metric, such as correct field extraction or appropriate tool selection.
- 02Choose the variant whose inputs and outputs match the application; do not assume one variant inherits every modality from another.
- 03Prepare representative data and set aside examples that are not used for fine-tuning.
- 04Compare the base and fine-tuned models on the same cases, record errors, and also evaluate data outside the training sample.
- 05Measure latency, memory use, and cost on the target hardware before integrating the model into a production workflow.
Availability, inference, and practical requirements
Liquid AI says LFM3 models have open weights and are available in the Hugging Face model catalog. The announcement includes examples using Transformers and installation commands for text, vision, and audio dependencies. Running each variant requires libraries and components appropriate to its modality; for example, the LFM3-VL example installs PyTorch, torchvision, and Pillow, while the audio examples add tools for processing sound.
“Open weights” alone do not spell out all legal and operational conditions of use. The information in the announcement is not enough to summarize a single license that applies to every artifact, or to confirm that the conditions are identical across models. Check the card and terms for the specific variant before incorporating it into a product or redistributing it.
The company says LFM3-1.7B can take about 1 GB at four-bit precision and that the text model supports a context of up to 256,000 tokens. These are figures Liquid AI reports for a particular configuration: download size, memory during inference, and speed also depend on the weight format, inference engine, hardware, and input length. They are not guarantees that the model will run on every device.
What to check before deployment
Advertised features can help narrow the options, but they do not replace testing in the intended environment.
| Aspect | What the announcement says | What to evaluate in the application |
|---|---|---|
| Hardware | The company says its models can be fine-tuned on one GPU. | Memory available, fine-tuning time, and inference performance on the chosen machine. |
| Quantization | An approximate 1 GB size is cited for the text model at four-bit precision. | Quality after quantization, actual memory use, and compatibility with the inference engine. |
| Context | A context window of up to 256,000 tokens is announced for the text model. | Quality and latency at the context length the application needs. |
| License | The models are described as having open weights. | Terms specific to the artifact and intended use. |
How to interpret the performance figures
The announcement compares LFM3 with earlier models and some third-party systems on agent, vision, audio, and decision tests. Among other figures, Liquid AI says LFM3-1.7B outperforms LFM2.5-2.6B on certain tasks, and that LFM3-VL-2B raises the average across 62 vision tests compared with its predecessor. The company also reports speed and speech-quality results for specific components.
These metrics are results attributed to the company’s own evaluations. The available documentation does not establish independent replication and is not enough to determine how each model will perform on different data, languages, hardware, or workflows. Comparisons can only be interpreted alongside the test definition, dataset, configuration, and scoring criteria; a score is not a guarantee of general performance.
The primary source is a Liquid AI launch post documenting the company’s claims, described architecture, and usage examples. Hugging Face’s article index confirms that the announcement appeared on October 27, 2026, but it does not add independent technical validation. The most supportable reading, therefore, is that LFM3 is a family and an adaptation path announced by its developer; its value for a particular application must be established through application-specific tests under comparable conditions.
Open questions
- The verified sources do not include an independent evaluation confirming the performance metrics reported by Liquid AI.
- The available information does not establish one license for all LFM3 models and artifacts; terms may vary by model card.
- Performance, memory use, and speed in a particular application cannot be inferred solely from the advertised parameter count, quantized size, or context window.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction