Liquid AI releases open-weight LFM2.5-24B-A2B, a 24B MoE reasoning model with 2B active parameters for on-device agents

Verified (fact)

Liquid AI unveiled 'LFM2.5-24B-A2B' on the Hugging Face blog. It is an MoE reasoning model with 24B total parameters and 2B active parameters per token, targeting private agents that run on-device. It always thinks before answering. It is the largest and final model in the LFM2.5 family, which started at 230M. A GGUF version and a free demo were also released.

Needs fact-check (claim)

Liquid AI claims it matches Qwen3.6-35B-A3B, which has about 50% more total and active parameters, on agent and tool-use tasks. It reported a LiveClaw score of 48.09, higher than Qwen3.6-35B-A3B's 38.19.

Why it matters (interpretation)

The claim is that an MoE with around 2B active parameters can boost tool-use performance for local agents. This widens the options for building on-device agents that keep data from leaving the device.

Sources

• Official Run fast, private agents on-device with LFM2.5-24B-A2B Hugging Face
• Community LiquidAI/LFM2.5-24B-A2B Reddit

More about Liquid AI

• 2026-10-07 P2 Open source Open weights MultimodalEfficiency & infra Liquid AIHugging Face
  Liquid AI unveils d1, multimodal open 'decision models' for edge devices

  Liquid AI posted on the Hugging Face blog to introduce 'd1', multimodal open decision models for edge devices. Community members...