8.3B total / 1.5B active (MoE)
128,000 tokens
LFM 1.0 license
Official open-weight LFM2.5 family release covering a pre-trained base checkpoint and a reasoning-tuned general-purpose checkpoint, positioned by Liquid AI as an on-device personal-assistant model line for tool use and complex instruction following on consumer hardware. Release materials say LFM2.5-8B-A1B builds on the earlier LFM2-8B-A1B with scaled pretraining from 12T to 38T tokens, large-scale reinforcement learning, an expanded context window from 32,768 to 128,000 tokens, and a doubled 128K vocabulary for better non-Latin tokenization efficiency. The published config identifies Liquid's hybrid sparse Lfm2MoeForCausalLM architecture with 24 layers, 32 experts, 4 experts active per token, 32 attention heads, 8 KV heads, 18 double-gated LIV convolution layers plus 6 GQA attention layers, and 128,000 maximum position embeddings. The Hugging Face card reports strong gains over LFM2-8B-A1B across instruction following, math, function calling, and hallucination-sensitive evaluations, including IFEval, IFBench, Multi-IF, MATH500, AIME25, BFCLv3, BFCLv4, and AA-Omniscience metrics. Liquid also emphasizes day-one deployment support across Transformers, llama.cpp, MLX, vLLM, SGLang, and ONNX, with the tuned checkpoint specifically recommended for agentic workflows, structured outputs, multilingual assistants, and on-device personal-assistant applications.