Liquid AI dropped a reasoning model on January 20 that runs entirely on-device. LFM2.5-1.2B-Thinking generates internal chains of thought before answering, the same approach used by larger reasoning models, but squeezed into 1.2 billion parameters.
The model shows big jumps over its non-thinking sibling. Math reasoning climbed from 63 to 88 on MATH-500, instruction following went from 61 to 69 on Multi-IF, and tool use improved from 49 to 57 on BFCLv3. These are Liquid AI's own benchmarks, though the company reports they follow ArtificialAnalysis methodology.
Liquid claims the model matches or beats Qwen3-1.7B on most reasoning tasks despite having 40% fewer parameters. The efficiency angle is interesting: fewer output tokens for comparable results, which matters when you're running on battery.
New launch partners include Qualcomm, Ollama, FastFlowLM, and Cactus Compute, joining existing partners AMD and Nexa AI. The company reports 239 tok/s decode on AMD CPU and 82 tok/s on mobile NPU. Memory footprint stays under 1GB.
Weights are available now on Hugging Face, LEAP, and Liquid's playground. The technical report covers the training recipe, which uses curriculum RL and iterative model merging across 25 checkpoints.
The Bottom Line: A sub-1GB reasoning model that runs offline on phones, with benchmark scores that put pressure on larger competitors.
QUICK FACTS
- Model size: 1.2B parameters, under 900MB memory
- MATH-500 score: 88 (company-reported)
- Inference speed: 239 tok/s decode on AMD CPU, 82 tok/s on mobile NPU
- Launch partners: Qualcomm, AMD, Nexa AI, Ollama, FastFlowLM, Cactus Compute
- License: Open-weight
