Text, Voice, Video, and Decisions: Liquid AI Brings Together Six Compact Models in LFM3

Liquid AI unveils six specialized models under 3 billion parameters, capable of processing text, image, video, and audio directly on devices.

A 1.9-billion-parameter vision model that competes with systems six times its size, local voice assistants, and specialized AI that can be fine-tuned on a single GPU. Liquid AI's LFM3 family consists of six models designed to process text, images, video, audio, and structured decisions without necessarily relying on cloud infrastructure.

The third generation of Liquid Foundation Models shares a common architecture pretrained on 22 trillion tokens. Its individual variants address different needs, including personal assistants, document analysis, robotics, and multimodal processing. Model weights are available through Hugging Face, alongside tools for customization and deployment on personal and enterprise hardware. A shared architecture designed around hardware constraints

Liquid AI is pursuing a different direction from laboratories focused primarily on building ever-larger models. Its approach aims to increase capability per parameter while accounting for memory consumption, processing power, and the devices expected to run the models.

LFM3 introduces a tri-hybrid architecture combining three mechanisms: gated short convolutions, sliding-window attention, and global attention.

The base model contains 24 layers, including 13 convolutional layers, five sliding-window attention layers, and six global attention layers. The combination is designed to keep inference efficient while supporting longer and more complex tasks.

Liquid AI used an automated hardware-in-the-loop architecture search process, evaluating model quality alongside measured inference speed and memory usage.

The search covered 14 hardware and backend configurations, including smartphones, Apple Silicon computers, AMD Ryzen AI processors, a Raspberry Pi 5, and Qualcomm Hexagon accelerators.

The resulting models share a foundation but incorporate different components depending on the type of data they process. LFM3-1.7B fits its quantized weights into approximately 1 GB

The text model, LFM3-1.7B, contains approximately 1.7 billion parameters and supports a context window of up to 256,000 tokens.

At 4-bit precision, its weights occupy approximately 1 GB of memory, making local deployment possible on compatible personal computers and mobile devices.

The extended context is particularly relevant to agents that must retain long conversation histories, process multiple documents, and execute tasks involving numerous tool interactions.

According to Liquid AI's evaluations, LFM3-1.7B outperforms its predecessor, LFM2.5-2.6B, while using 35% fewer parameters.

On WebArena-Verified and AutomationBench, two benchmarks evaluating multistep tasks across browsers and business applications, it also outperforms Qwen3.5-9B under the company's testing conditions.

Memory efficiency remains important as conversations and agent histories grow. At a 128,000-token context length, Liquid AI reports a 1.6 GB KV cache, compared with 4.3 GB for Qwen3.5-4B.

To demonstrate large-scale execution, the company also ran more than 300,000 LFM3-1.7B agents concurrently to investigate the Enron email archive. The experiment demonstrates large-scale orchestration rather than independently establishing the accuracy of every agent's findings. A 1.9B vision model competes with much larger systems

LFM3-VL-2B combines text, image, and video understanding. With approximately 1.9 billion parameters, it can analyze documents, interpret charts, describe scenes, and reason about the spatial relationships between objects.

Across 62 visual evaluations, the model achieves an average score of 71.4. Liquid AI's detailed results place Gemma 4 12B at 71.7, a comparable score from a model with approximately six times as many parameters.

Performance varies by category. LFM3-VL-2B scores 83.1 on graphical user interface understanding, compared with 76.1 for Gemma 4 12B. It also scores 77.3 on visual grounding, versus 74.1.

The comparison reverses for mathematical and logical reasoning, where Gemma 4 12B scores 79.3 against LFM3-VL-2B's 68.8. Liquid's compact model also trails more capable competitors on some general video and agentic tasks.

Its new in-house vision encoder uses approximately three times less compute than the SigLIP2 encoder in previous Liquid vision models. For video, the encoder compresses each pair of sampled frames into one temporal representation, reducing the sequence length the model must process.

These optimizations support applications such as analyzing videos recorded on phones, understanding technical documents, and performing local visual inspection. Voice conversations across 18 languages

LFM3-Audio-2B adds speech understanding and generation to the family. It accepts text or audio and can respond through text, speech, or both.

The model supports 18 languages and handles transcription, translation, audio understanding, and spoken conversations. It can also convert voice instructions into tool calls that interact with software and devices.

Its causal audio encoder begins processing microphone input while the user is speaking. Liquid AI reports time to first token as low as 50 milliseconds under its measured conditions.

Speech generation uses LiquidCodec, a tokenizer that converts generated tokens into audio. On a single Apple M4 Max CPU thread, Liquid AI reports decoding at approximately 87 times real-time speed. That figure concerns audio decoding, not the total response latency of a voice assistant.

On Liquid AI's internal spoken tool-calling evaluation, LFM3-Audio-2B scores 58.5, compared with 44.3 for Qwen2.5-Omni-3B. On VoiceBench, it scores 73.1, ahead of Gemma 4 E2B at 63.3, but below MiMo-Audio-7B-Instruct at 77.3.

The results demonstrate strengths in specific tasks rather than overall superiority across speech benchmarks. Combining vision, sound, and language in one model

LFM3-Omni-2B brings visual understanding, audio processing, and language together in a 2.2-billion-parameter model.

It can process a video alongside its soundtrack and answer questions that require