GoKawiilLiquid AI launches LFM2.5-24B-A2B reasoning model for on-device agents
Liquid AI released LFM2.5-24B-A2B, a 24-billion-parameter mixture-of-experts model with only 2 billion active parameters per token, as the largest and final entry in its LFM2.5 family. The model is reasoning-only, generating an explicit chain of thought, and both base and post-trained open-weight versions are now available on Hugging Face and Liquid's Playground.
GoKawiil's interpretation of the reporting above, not reported fact.
By activating far fewer parameters than its total size, the model can run quickly on consumer hardware like an Apple M5 Max while reportedly matching or beating a larger rival, Qwen3.6-35B-A3B, on several agentic benchmarks. This suggests Liquid AI is prioritizing efficient, private, on-device agent deployment over simply scaling parameter counts, a strategy that could appeal to developers wanting local inference without cloud dependency.
- LFM2.5-24B-A2B has 24B total but only 2B active parameters per token, enabling faster on-device inference.
- It outperformed Qwen3.6-35B-A3B on 5 of 8 agentic and tool-use benchmarks despite fewer active/total parameters.
- Both base and post-trained versions are open-weight and available now on Hugging Face and Liquid AI's Playground.
Apple Mac mini M4 — This article highlights on-device AI performance on Apple Silicon like the M5 Max, and the Mac mini M4 is an affordable way to get into that same Apple Silicon ecosystem for running local LLMs. Its unified memory architecture is great for experimenting with models like LFM2.5 directly on your desktop without relying on the cloud.
See Apple Mac mini M4 on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: liquid.ai, 2026-10-14
Published there as: “LFM2.5-24B-A2B: Built for fast on-device agents”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.