Liquid AI launches LFM3: Six open‑weight 2B‑class multimodal models

Liquid AI announced the LFM3 family, a collection of six open‑weight multimodal models that sit in the 2 billion‑parameter range. The launch is notable because the models are released with permissive licenses, covering text, vision, audio, omni‑modal and decision‑making tasks, and because the flagship 1.7 B variant packs a 256 K context window into roughly 1 GB of memory when quantised to 4‑bit. For creators, developers and researchers who rely on accessible, high‑capacity models, LFM3 adds a new tier of capability without the cost of proprietary large‑scale services.

What is LFM3 and why it matters?

According to the MarkTechPost report, LFM3 is positioned as an open‑weight alternative to the closed‑source multimodal giants that dominate the market. Each model in the suite is built to handle a specific modality or a combination of them, allowing users to pick the most suitable engine for their workflow. The open‑weight nature means the weights can be downloaded, fine‑tuned, or integrated into custom pipelines, fostering experimentation and reducing reliance on vendor‑locked APIs. In practice, this expands the toolbox for independent developers and academic labs that need large context windows and multimodal reasoning but lack the budget for commercial offerings.

Which modalities are covered and what are the six models?

The LFM3 family comprises six distinct models:

1. LFM3‑Text – optimized for long‑form language generation and understanding.
2. LFM3‑Vision – designed for image captioning, visual question answering, and feature extraction.
3. LFM3‑Audio – tuned for speech‑to‑text, audio classification, and music‑related tasks.
4. LFM3‑Omni – a hybrid that can ingest text, images, and audio in a single prompt.
5. LFM3‑Decision – a model focused on planning, policy selection, and structured decision problems.
6. LFM3‑1.7B – the flagship model that combines the three‑hybrid backbone approach and a 256 K token context window.

All models sit around the 2 billion‑parameter mark, placing them in the same class as many current open‑source LLMs but with added multimodal heads. The release includes the model weights, tokenizer files, and inference scripts, making it straightforward to spin up a local server or a cloud instance.

How does the tri‑hybrid backbone and 256K context differentiate LFM3‑1.7B?

The flagship LFM3‑1.7B model uses a "tri‑hybrid" architecture that blends three proven design patterns: a transformer‑based language core, a vision‑specific convolutional stem, and an audio‑oriented spectrogram encoder. By merging these components, the model can process heterogeneous inputs without requiring separate preprocessing pipelines. The 256 K context window—ten times larger than the typical 8‑16 K windows of most 2 B‑class models—means the model can retain far more information from a single prompt, which is especially valuable for long documents, code bases, or video transcripts.

Quantising the model to 4‑bit reduces the memory footprint to roughly 1 GB, enabling it to run on consumer‑grade GPUs (e.g., RTX 3060) or even high‑end CPUs with modest latency. This combination of breadth (multimodal) and depth (long context) is rare in the open‑weight space and could open new use‑cases such as multi‑hour podcast summarisation, document‑level reasoning, or interactive design assistants that remember extensive visual histories.

Practical implications for developers and creators

For developers, the open‑weight nature removes the API‑call cost barrier. You can host LFM3 on‑premises, embed it in web services, or integrate it into existing pipelines. The long context window reduces the need for chunking strategies, simplifying code that deals with large texts or multimodal streams. Creators can experiment with cross‑modal prompts—e.g., feeding an image and a short audio clip to generate a descriptive paragraph—without stitching together separate models.

Because the models are released under permissive licenses, fine‑tuning is viable. Teams can adapt LFM3‑Vision for a niche visual domain (medical imaging, fashion) or train LFM3‑Audio on a specific dialect. The relatively small size (1 GB at 4‑bit) also means fine‑tuning can be performed on a single GPU, lowering the entry threshold for small studios and research groups.

How to try LFM3 with Grafy

Grafy users can explore LFM3 without leaving the platform. The Grafy Chat interface lets you load any compatible model from a URL or local path and converse with it alongside other models, making side‑by‑side comparisons trivial. To experiment with the multimodal capabilities, upload an image or audio file in the chat window and prepend a prompt that references the other modality; the tri‑hybrid backbone will handle the mixed input automatically.

If you prefer to write code, the Grafy Code studio provides a ready‑made Python environment where you can install the LFM3 inference package, load the 4‑bit weights, and run inference scripts. Sample notebooks are included in the release archive, and Grafy Code can execute them directly in the browser, giving you instant feedback on latency and memory usage.

For those who simply want to read the original announcement or dive deeper into the technical report, the Grafy Reader app can ingest the MarkTechPost article (or the PDF of the release notes) and let you ask follow‑up questions, search for specific sections, or extract tables—all powered by the same LFM3 models you are evaluating.

Outlook for open‑weight multimodal research

The introduction of LFM3 adds a competitive pressure on closed‑source providers by showing that sizable multimodal capability can be achieved at the 2 B‑parameter scale. Researchers may use the publicly available weights as baselines for new architectural experiments, such as more efficient token‑reduction strategies for ultra‑long contexts or tighter integration of audio‑visual attention mechanisms.

The release also signals a broader trend: open‑weight projects are moving beyond pure language models toward unified, cross‑modal systems. As more groups adopt LFM3 and contribute community fine‑tunes, a virtuous cycle of improvement could emerge, lowering the barrier for applications like AI‑driven video editing, interactive storytelling, and real‑time decision support.

---

Frequently asked questions

What hardware is required to run LFM3‑1.7B?

When quantised to 4‑bit, the model occupies about 1 GB of VRAM. A modern consumer GPU (e.g., NVIDIA RTX 3060 or higher) or a CPU with sufficient RAM can host the model for inference, though GPU acceleration will give lower latency for multimodal inputs.

Are the LFM3 weights truly open‑weight?

Yes. Liquid AI released the weights under a permissive license that allows downloading, redistribution, fine‑tuning, and commercial use, as stated in the MarkTechPost release.

How does Grafy Chat handle multimodal inputs?

Grafy Chat accepts file uploads (images, audio) alongside text prompts. When you load an LFM3 model that supports the corresponding modality, the backend routes each input to the appropriate encoder in the tri‑hybrid backbone, enabling seamless cross‑modal interactions.

---

Source: MarkTechPost — Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models