Liquid AI Releases LFM3: Six 2B-Class Open-Source Multimodal Models

Liquid AI Unveils Another Small-Model Matrix: LFM3 Drops Six 2B-Class Open-Weight Models at Once

Liquid AI has released the LFM3 series, launching six open-weight models in a single batch, all positioned at the 2B parameter scale and spanning text, vision, audio, omni, and decisions directions. The move continues the company's strategy of betting on "small size, high usability" models.

One Release, Five Modalities

Based on official disclosures, LFM3 is not a single model but a model family: the six models each correspond to different input-output configurations. Text, vision, and audio are the three capability lines most commonly combined in current multimodal products, while "omni" means a single model handling multiple modalities simultaneously, and "decisions" points to decision-making tasks—the latter being relatively rare among open-source small models, making its specific task definitions and evaluation methods worth watching.

Bringing multimodal capabilities down to the 2B level is the core product logic of this model set. Smaller parameter counts mean lower inference costs and lower deployment barriers, making them suitable for on-device or edge scenarios; however, multimodal capabilities at small parameter counts often come with capability degradation, and actual performance will need to await independent evaluation.

"Open weights" and "open source software" are not the same concept: the former typically only releases the trained model parameters, without disclosing training data, training code, or the complete recipe. Therefore, users' rights depend entirely on the license terms—whether commercial use is permitted, whether redistribution of fine-tuned models is allowed, and whether attribution or open derivatives are required. For teams planning to embed models into products, the license is often more decisive than benchmark scores.

"Omni models" refer to a single model that simultaneously accepts and generates multiple modalities such as text, images, and audio, rather than a pipeline assembled from multiple specialized models. The advantage is that cross-modal information can interact within the same parameter space; the disadvantage is that capabilities across modalities may compete with each other, and it becomes difficult to finely optimize for a single modality. This is precisely the key to judging how to choose between omni and specialized text/vision/audio models.

LFM3-1.7B: Tri-Hybrid Backbone and 256K Context

Officials highlighted the LFM3-1.7B model. It uses a tri-hybrid backbone structure, supports a 256K context window, and is approximately 1GB in size at 4-bit quantization.

These three points together point to the same goal: preserving long-context capability within an extremely small footprint. A 256K context is uncommon among 2B-class models, and "approximately 1GB at 4-bit" means it can fit into memory-constrained devices. The tri-hybrid backbone is Liquid AI's technical signature; which types of layers it specifically mixes and what role each plays was not elaborated in official materials, but from the naming it should be a combination of different sequence modeling mechanisms to balance efficiency and expressive capacity.

Hybrid backbones have been an active direction in sequence modeling in recent years. The basic idea is to combine different types of layers within the same network—common combinations include standard attention layers, linear attention/state space model (SSM) layers, and convolutional or gated linear unit layers. Pure Transformer attention computation grows quadratically with sequence length, making long contexts expensive; linear attention or SSM layers can process long sequences with near-linear complexity, but often fall short of full attention in precise recall and complex reasoning. Hybrid architectures balance efficiency and expressive capacity by having most layers take the cheap path while a small number retain full attention. The industry already has practices such as Mamba-2/Transformer hybrids and Jamba.

4-bit quantization refers to compressing model weights from 16-bit floating point to 4-bit integer or low-bit floating point representations, which can reduce memory usage to about one quarter—a key technique for on-device deployment. However, quantization introduces precision loss, which may amplify errors especially in long-context and multimodal tasks. Therefore, "1GB at 4-bit" is only a size metric and does not equate to a quality guarantee at that precision.

Significance for the Open-Source Small Model Space

Open weights + small parameters + multimodality—these three conditions combined directly target a group of teams working to compress multimodal capabilities onto end devices. For developers, what truly matters is not just "it runs," but license terms, fine-tuning friendliness, and actual quality across modalities.

Currently available information focuses on model specifications: six models, 2B class, the backbone structure of the 1.7B model, context length, and quantized size. As for training data composition, benchmark results, and license details, the original materials do not cover them—yet these are precisely the key factors determining whether a model can be adopted in production environments.

On-device deployment refers to models running directly on phones, laptops, embedded devices, or edge servers rather than calling cloud APIs. The constraints are not just memory: they also include power budgets, first-token latency, heat generation during sustained inference, and privacy compliance requirements. A 2B-class model at approximately 1GB after 4-bit quantization can theoretically fit into the memory of mid-to-high-end mobile devices, but the KV cache required for 256K context grows linearly with sequence length, and actual memory usage in long-context scenarios is far higher than the weight size—an easily overlooked point when evaluating "whether true on-device long context is feasible."

Questions to Watch

First, whether the capability boundaries of the six models are clear—especially whether there is capability overlap and trade-off between "omni" and the separate text/vision/audio models. Second, the task scope of the decision-making model needs clear definition, otherwise its practical value is hard to assess. Third, the actual effective length of 256K context on a 1.7B model typically falls short of the nominal value.

The LFM3 release shows that the multimodal race among small models is still accelerating. The specifications are now on the table; next comes evaluation data and real-world deployment feedback.