Liquid AI's LFM2.5-230M Runs at 213 tok/s on a Phone CPU

LFM2.5-230M is Liquid AI's smallest model to date: a 230-million-parameter open-weight model built specifically for on-device agentic workloads. The pitch is deliberately narrow. This is not a general-purpose assistant or a reasoning engine. It is a fast, deployable agent layer for data extraction and tool use, designed to run on hardware that most models would choke on.

A different kind of small model

The small-model space in 2026 has split into two camps: reasoning specialists that compress math and coding into compact parameter counts, and edge agents that prioritize throughput and deployability over raw intelligence. LFM2.5-230M is Liquid AI's bet that the next wave of useful AI is a 230M-parameter open-weight agent that runs at 213 tok/s on a phone CPU, parses natural language into tool calls on a humanoid robot, and beats models twice its size on instruction following and data extraction.

This release distinguishes itself from other small models by utilizing the LFM2 architecture to achieve high inference speeds without the massive memory overhead typical of parameter-heavy transformers. The architecture is a hybrid: 14 layers total, with eight double-gated LIV convolution blocks plus six grouped-query attention blocks, targeting fast CPU inference, a 32,768-token context, and ten languages. The convolution blocks are the key differentiator here. Unlike pure-attention transformers, they process sequences more efficiently on constrained hardware by avoiding the quadratic cost of full attention at every layer.

How it was built

The model was pre-trained on 19 trillion tokens, including a 32K context extension phase. After pre-training, the team applied a three-stage post-training recipe:

1. Supervised fine-tuning with distillation from the larger LFM2.5-350M -- this is what lets a 230M model punch above its weight; it inherits behavioral patterns from the bigger sibling on targeted tasks.