LFM2.5-230M

LFM2.5-230M, published on 24 June 2026, is the smallest member of Liquid AI's LFM2.5 family and the company's answer to hardware where even a two-billion-parameter model will not fit. It has 229.7 million parameters in 14 layers — 8 double-gated convolution blocks and 6 grouped-query attention blocks — a 32,768-token context and a 65,536-token vocabulary, half the vocabulary of its 2.6-billion-parameter sibling. The model was not trained from scratch to this size in the usual sense: it is distilled from LFM2.5-350M and then refined through multi-stage reinforcement learning, on a pre-training budget of 19 trillion tokens. Liquid AI positions it for tool use and data extraction rather than open-ended conversation, and the knowledge cutoff is mid-2024 — older than the model itself by two years, which matters for anything that depends on recent facts. The numbers the company reports are about hardware rather than benchmarks: 213 tokens per second of decoding on a Galaxy S25 Ultra and 42 tokens per second on a Raspberry Pi 5. That places it in a different bracket from most catalogue entries — this is a model for a sensor, a handset or a single-board computer, not for a server. Ten languages are declared, and Polish is not among them, unlike in the larger LFM2.5-2.6B. Weights are downloadable in native, GGUF, ONNX and MLX builds, with a separate base checkpoint for fine-tuning. The licence is the same LFM Open License v1.0 that governs the rest of the family: free for research and redistribution, but commercial use is licensed only below ten million dollars of annual revenue.

▸Videos

No videos yet.

▸Reviews