A small model aimed at local agents
Liquid AI released LFM2.5-2.6B on Hugging Face on August 4. Its model card lists 2.69 billion parameters, 30 layers, a maximum 131,072-token context, 16 languages including Korean and Japanese, and support for tool use and multi-step agent tasks. It lists 34 trillion training tokens.
Day-one integrations include llama.cpp, MLX, vLLM, SGLang and ONNX. Liquid AI reports 220 tokens per second on an M5 Max, 113 on a Ryzen AI Max+ 395, under 2.5 GB of memory, and roughly 30 tokens per second on a phone. These are vendor measurements; hardware, quantization and input length can materially change results.
Open weights are not automatically open source
The weights are downloadable, but the card uses the dedicated LFM Open License v1.0 and marks the license as “other,” rather than Apache or MIT. It should therefore not be described automatically as OSI open source. Teams must read the original terms for commercial use, redistribution, derivatives and restricted uses.
Fit and limitations
The model is interesting for local document retrieval, privacy-sensitive tool execution and repeatable tasks on lower-power hardware. Liquid AI also acknowledges that coding can trail larger models. Public benchmarks and vendor speed figures do not guarantee production quality; teams should reproduce task success, memory use and error behavior with their own languages and tool schemas.