Efficient foundation models using a hybrid backbone combining gated short convolutions with grouped query attention. 350M to 24B parameters (MoE, 2.3B active). Optimized for edge and on-device deployment.

LFM2-24B achieves 239 tok/s (#3 speed ranking on AA). Also includes vision-language (LFM2.5-VL), speech, and pharmaceutical variants (with Insilico Medicine). AA Intelligence Index: 5 (24B). Apache 2.0.

Model Details

Architecture MOE
Parameters 24B
Active params 2.3B
AA Intelligence 5

Variants

Name Parameters Notes
LFM2-350M 350M
LFM2-1.2B 1.2B
LFM2-2.6B 2.6B
LFM2-8B-A1B 8.3B
LFM2-24B-A2B 23.8B
LFM2.5-350M 350M Released April 2026, trained on 28T tokens with RL, <500MB footprint
LFM2.5-8B-A1B 8.3B Released May 28 2026; hybrid MoE on-device flagship, 128K context, reasoning-only with explicit CoT, scaled to 38T training tokens. IFEval +12.4, AIME25 +22.5 vs LFM2-8B-A1B. AA Intelligence Index 8 (AA v4.1) — above LFM2-24B-A2B's 5. GGUF/MLX/ONNX variants shipped.
LFM2.5-1.2B-JP / Audio-1.5B-JP 1.2B Japanese-language variants released June 4 2026 (text + audio).
LFM2.5-VL-Extract (450M / 1.6B) 1.6B Vision-language Extract variants for OCR / structured document extraction, released June 4 2026 (450M and 1.6B sizes, GGUF available).
LFM2.5-230M 230M Released June 24 2026; Liquid's most compact model yet — hybrid 8 double-gated LIV-conv + 6 GQA blocks, 32K context, 19T tokens, distilled from LFM2.5-350M. IFEval 71.71. LFM Open License v1.0. ~30K downloads in first week.
LFM2.5-Embedding-350M / ColBERT-350M 350M First retrieval members of the family (announced June 18 2026): bidirectional-mask 350M retrievers on LFM2.5-350M-Base, 11 languages; NanoBEIR NDCG@10 0.605 / 0.577.

Paper

open-weighton-devicemoe

Related