Reproduce “LFM2 Technical Report” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexander Amini, Anna Banaszak, Harold Benoit, Arthur Böök, Tarek Dakhran, Song Duong, Alfred Eng, Fernando Fernandes and 25 more
LFM2 is a family of models designed for local assistants, using gated short convolutions in most layers and a few grouped-query attention blocks for longer-range context. On a Samsung Galaxy S25, LFM2-350M reached 657 tokens per second on 4K-token prompt prefill, versus Granite-4.0-350M’s 210; these are batch-one CPU throughput measurements, not universal speed or energy results.
Key insights
- A "minimal hybrid architecture" tailored through hardware-in-the-loop search can outperform more complex sub-quadratic or hybrid architectures for on-device efficiency, achieving superior quality-efficiency trade-offs on CPUs.
- Sophisticated pre-training (e.g., decoupled Top-K knowledge distillation, extended context length training) and post-training (e.g., detailed SFT data curation, curriculum learning, advanced preference alignment) techniques are crucial for maximizing the capabilities of smaller, efficient models.
- The "edge-first" philosophy can be successfully extended to multimodal (vision-language, speech) and information retrieval tasks, enabling real-time, privacy-preserving, and multilingual AI applications on commodity hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.