Liquid AI publishes LFM2 technical report
Liquid AI has released the full technical report for LFM2, its second-generation foundation models for edge devices, documenting architecture search, distillation, and post-training methods alongside open weights.
Liquid AI announced the release of the full technical report for LFM2, the second generation of its Liquid Foundation Models. The company builds efficient models designed to run on edge devices, and says it is publishing the report so its methods are not kept proprietary.
According to the company, the report, which is posted on arXiv, covers:
- Hardware-in-the-loop architecture search
- New training objectives for knowledge distillation
- Post-training recipes for small models
- Vision-language capabilities (LFM2-VL)
- Speech processing (LFM2-Audio)
- Information retrieval (LFM2-ColBERT)
All LFM2 models are available with open weights on Hugging Face. They come with deployment guides for ExecuTorch, llama.cpp, and vLLM, which cover on-device, local CPU/GPU, and server inference.
The blog post gives no benchmark results, model sizes, or license terms. It also does not describe the architecture or the distillation objectives in any detail. Those details are presumably in the arXiv paper itself.
For teams building small models, the most useful parts of the report will likely be the hardware-in-the-loop search and the post-training recipes for small models. Methods like these are often left out of open-weight releases.