Liquid AI releases LFM3 multimodal model family for on-device specialized intelligence

Liquid AI launched LFM3, a multimodal model family integrating text, vision, audio, and decision capabilities into a single shared backbone for on-device and edge deployment.
The 1.7B text model achieves a 49.37 average benchmark score, outperforming Qwen3.5-9B (47.24) and LFM2.5-2.6B (31.55), while using 1.6 GB of KV cache for 128K context compared to 4.3 GB for Qwen3.5-4B.
LFM3-VL-2B matches the performance of Gemma 4 12B (6x larger) with a 71.4 average score, and LFM3-Audio-2B supports 18 languages with 6.2% English WER and 50 ms time-to-first-token.
New decision models (d1-2B and d1-Omni-2B) enable millisecond-scale responses via single forward passes, answering questions with calibrated probabilities.

Architecture and Training

Tri-hybrid layer mix combines 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers to balance speed and quality.
Pre-trained on 22T tokens with a 256K context window, including increased data for scientific papers, code, and agentic traces.
Post-training involves SFT, mixed RL, multi-domain on-policy distillation, and agentic RL to optimize for tool use and specialized workflows.

Open-weight models are available on Hugging Face and the Liquid AI API, with native support for llama.cpp, MLX, and vLLM.
Launch partners include AMD, Arm, Intel, NVIDIA, Qualcomm, and Vercel, with pre-release testing by Mercedes-Benz and Shopify.
Hardware support spans 14 device/backend combinations, including Samsung Galaxy S26 Ultra, Raspberry Pi 5, and Apple M4 Max.