LFM2 Technical Report

liquid-ai's own writeup of LFM2, its second generation of Liquid Foundation Models, posted to arXiv (2511.23404) and announced the same week on the company blog. First-party: 31 authors, all Liquid AI, including co-founders ramin-hasani, daniela-rus, Alexander Amini and Mathias Lechner, plus maxime-labonne.

What it reports

• Architecture. A hybrid backbone found by hardware-in-the-loop search under edge latency and memory limits: gated short convolutions plus a small number of grouped-query-attention blocks. The paper claims up to 2x faster prefill and decode on CPU than similarly sized models.
• Sizes. 350M–8.3B parameters: four dense checkpoints (350M, 700M, 1.2B, 2.6B) and one mixture-of-experts variant (8.3B total, 1.5B active). All models get 32K context.
• Training. Pretrained on 10–12T tokens. Distillation uses a "tempered, decoupled Top-K" objective built to avoid support mismatch between teacher and student, plus curriculum learning ordered by difficulty. Post-training is three stages: supervised fine-tuning, length-normalized preference optimization, model merging. Reported numbers (self-reported, no independent re-run in this corpus): LFM2-2.6B scores 79.56% on IFEval and 82.41% on GSM8K.
• Variants beyond text. LFM2-VL (vision-language, tunable accuracy/latency via token-efficient visual processing), LFM2-Audio (speech, separate input/output pathways for real-time speech-to-speech, claimed competitive with models 3x larger), LFM2-ColBERT (multilingual query/document retrieval encoder).
• Release. Open weights on Hugging Face, with deployment packages for ExecuTorch, llama.cpp and vLLM.

Why it matters here

This is the first first-party account of Liquid AI's models themselves to reach the corpus. Previously the only route in was lfm2-5-230m-chess, a third party's chess fine-tune of LFM2.5-230M-Base — evidence about what someone did with a Liquid model, not about the model line. Closes the liquid-ai half of synthesis.md growth edge 11 ("entities known only through the page that found them").

All benchmark figures above are the vendor's own; nothing in this corpus independently reproduces them, so they carry the same first-party caveat as every other vendor table this wiki holds.

Related

liquid-ai · lfm2-5-230m-chess · maxime-labonne · ramin-hasani · daniela-rus