Liquid AI has announced the release of its LFM3 model family, a suite of compact, multi-modal artificial intelligence models designed primarily for edge computing and local deployment on mobile devices and laptops. The release includes six distinct models: text, vision-language, audio, omni-modal, and decision-making variants.

Text and Agentic Capabilities

The LFM3-1.7B model is a text-only language model utilizing a new tri-hybrid architecture consisting of 24 layers, including 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers. According to company documentation, the model features a 128,000-token vocabulary and supports a context window of 256,000 tokens.

Designed for local agents and specialized applications, the 1.73B parameter model has been post-trained for agentic harnesses such as Hermes Agent, Pi, and OpenClaw. At 4-bit precision, the model requires approximately 1 GB of memory, making it suitable for hardware with limited resources. While the model supports 16 languages, including English, Chinese, Arabic, and several European languages, Liquid AI notes it is not recommended for complex multi-step reasoning or agentic coding due to its small scale.

Multi-modal and Audio Integration

The LFM3-VL-2B is a vision-language model designed for on-device visual understanding. It integrates the LFM3 language model with an in-house vision encoder to process text, images, and video. The model is capable of spatial and embodied reasoning, which the company suggests provides a foundation for robotics applications.

The LFM3-Audio-2B model is a compact audio model that accepts text or audio inputs and produces text or speech outputs. It utilizes an in-house audio encoder and the LiquidCodec audio tokenizer. The model supports multilingual speech across 18 languages and is capable of interpreting recordings to answer questions regarding sounds and events.

Liquid AI also introduced LFM3-Omni-2B, an end-to-end omni-modal model. This 2.2B parameter model is designed to replace separate ASR, LLM, and TTS models by processing text, images, video, and audio within a single framework.

Decision Models and Performance

The release includes two decision models: d1-Omni-2B and d1-2B. These models are designed to return calibrated answers in a single forward pass without generating tokens.

According to technical specifications, d1-Omni-2B achieves decision speeds of 5.8 ms on an NVIDIA H100 and 3.7 ms on an AMD MI325X. The d1-2B model, built on the vision-language architecture, reportedly leads open decision benchmarks, outperforming the d1-3B model despite being significantly smaller.

Early Observations and Licensing

Initial user feedback on community platforms like Reddit has been mixed. Some users have noted that while the models show promise for local fine-tuning, performance in complex tasks—such as OCR or nuanced video narrative description—may currently lag behind larger models like Qwen or Gemma. One user reported that the vision model occasionally failed to correctly identify basic human postures and struggled with specific character recognition in OCR tasks.

The models are released under the LFM Open License v1.0. While the models are intended for specialized domain adaptation via single-GPU fine-tuning, some users have expressed a preference for more permissive licenses such as Apache or MIT.