LFM2.5-8B-A1B

LFM2.5-8B-A1B, published on 28 May 2026, is the largest model Liquid AI has released in its LFM2.5 family and the company's attempt to put mixture-of-experts inference on a phone. The weights file holds 8.47 billion parameters across 24 layers with a hidden size of 2,048, a 128,000-token context and a 128,000-entry vocabulary. The routing is the point of the design. The model carries 32 experts and activates four of them per token, so most of the network stays idle on any given step. That is what makes an eight-billion-parameter model tractable on a handset: memory holds all the weights, but arithmetic touches a fraction of them. The name deserves a footnote. "A1B" reads as one billion active parameters, and much of the coverage repeats that figure. Liquid AI's own model card gives a different number: 8.3 billion total and 1.5 billion active. The gap is not rounding — it is half again as much computation per token as the name advertises, which matters for anyone budgeting battery or latency on a device. Ten languages are declared — English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, Portuguese and Italian. Polish is not among them, as in the smaller LFM2.5-230M. The company positions the model as an on-device assistant chaining tool calls and following multi-step instructions, and claims day-one support for llama.cpp, MLX, vLLM and SGLang. A separate base checkpoint is published for fine-tuning. The licence is LFM Open License v1.0, the same terms as the rest of the family: the weights download without a gate, but commercial use is permitted only below ten million dollars of annual revenue. Above that threshold the user must come to terms with Liquid AI — a condition that "open weights" alone does not convey.

▸Videos

No videos yet.

▸Reviews