Liquid AI releases LFM3-Omni for offline use

The author tests Liquid AI's LFM3-Omni, a 2.2-billion parameter multimodal model that handles listening, looking, and talking locally on an 8GB M2 MacBook Air under 3GB of memory. Running via Ollama and a community MLX demo app with Wi-Fi disabled, the model accurately recognized various physical objects through the webcam and delivered fast transcription and speech capabilities. However, testing revealed limitations in numerical reasoning due to the Omni model's inability to reason before answering. Open article →