Liquid AI Launches LFM2.5‑24B‑A2B: 24B‑Parameter Reasoning Model with Only 2B Active Weights for On‑Device Agents

Liquid AI has released LFM2.5‑24B‑A2B, a reasoning‑focused mixture‑of‑experts (MoE) model that keeps 24 billion total parameters but routes each token through only 2 billion active weights. The architecture retains 40 layers, 64 experts per MoE block with top‑4 routing, and a hybrid of gated short convolutions and grouped‑query attention.

The context window expands from 32 K to 131 K tokens and the tokenizer vocabulary doubles to 128 K, improving multilingual support. Pre‑training scales to roughly 33 trillion tokens, followed by targeted mid‑training on reasoning, math, code, and long‑context data, then a five‑stage post‑training pipeline that turns the base model into a tool‑calling agent.

Benchmarks show LFM2.5‑24B‑A2B matching or exceeding larger MoE peers such as Qwen3.6‑35B‑A3B on agentic tasks (LiveClaw 48.09 vs. 38.19) while delivering 184 tokens/s on an Apple M5 Max and up to 21 K output tokens/s on a single NVIDIA B300 GPU.

Why it matters for GPU / AI infrastructure

Because only a small fraction of parameters are active per inference step, the model dramatically reduces memory bandwidth and compute requirements. This makes continuous, multi‑step agent workloads viable on edge devices and allows data‑center GPUs to serve many more concurrent sessions, lowering total cost of ownership for AI‑native applications.