Liquid AI Releases On-Device Reasoning Model LFM2.5-24B-A2B

Liquid AI has announced LFM2.5-24B-A2B, a model intended for reasoning and local deployment. The naming indicates 24 billion total parameters with about two billion activated for a given token under the model’s sparse architecture. Liquid AI has published performance and speed claims, but those results are vendor-reported and depend on hardware, quantisation, software and the chosen workload.

What on-device reasoning can offer

Running a model locally can reduce the need to send every prompt to a hosted service, improve responsiveness when the model and hardware fit the task, and allow some use without a continuous internet connection. Those benefits depend on the implementation. A local model may still use external tools or services, and local storage, logs and connected applications can remain privacy risks if they are not configured carefully.

A model described as “on-device” does not necessarily run comfortably on every phone or laptop. Memory requirements, prompt length, context size, token throughput, quantisation and thermal limits affect usability. The model may be more appropriate for a workstation or selected high-memory device than for a typical handset.

How to read the benchmarks

A model’s total parameter count is not the same as the amount of memory used after quantisation, and active parameters alone do not determine total runtime demand. Benchmark comparisons are most meaningful when hardware, inference engine, quantisation, context length, batch size and quality metric are stated. A speed figure in a short prompt test may not predict performance in a long reasoning task.

Teams considering local inference should also test factual accuracy, tool-use reliability, refusal behaviour, multilingual quality and the ability to recover from errors. For privacy-sensitive work, verify what is stored on the machine and whether telemetry or third-party extensions transmit data.

Who should evaluate it

LFM2.5-24B-A2B may be worth testing for developers seeking control over deployment and inference. The practical choice should be based on a representative local benchmark and total operating needs, not on a parameter-count label or a single vendor speed claim. Confirm licence terms, supported hardware and the release’s technical documentation before integrating it into a production application.