Liquid AI ships LFM2.5, a 2.6B on-device model

Liquid AI has released LFM2.5-2.6B, a 2.6-billion-parameter language model built for efficient, multilingual text generation on local hardware. The weights are published in GGUF format on Hugging Face, the quantization-friendly container used by runtimes such as llama.cpp and Ollama.

The pitch here is on-device inference. At 2.6B parameters, the model is small enough to run on a modern laptop or a capable phone without a cloud round-trip, which matters for latency, cost, and privacy. Shipping directly as GGUF signals that Liquid AI expects most people to run this locally rather than behind an API.

Why it matters

The compact-model tier has become one of the most competitive corners of open-weights AI, and Liquid AI is one of the more distinctive players in it thanks to its Liquid Foundation Model architecture.

• Small footprint: a 2.6B dense model fits comfortably on consumer hardware.
• Multilingual: the release is positioned for use across languages, not English alone.
• Local-first packaging: GGUF makes it straightforward to drop into existing on-device toolchains.

Some details, including context length and full benchmark numbers, aren't specified in the release record, and the model ships under a custom license rather than a standard open-source one — worth checking before commercial use. For teams building assistants that need to run offline or at the edge, LFM2.5-2.6B is a low-friction option to evaluate.