LiquidAI announced LFM2.5-24B-A2B, a 24-billion-parameter reasoning model engineered to run autonomous AI agents efficiently on consumer hardware. The model uses a Mixture of Experts architecture that activates only 2 billion parameters per token, drastically reducing computational requirements while maintaining competitive performance against much larger competitors. In benchmark comparisons, LFM2.5-24B-A2B outperformed models with roughly 50 percent more parameters on agent and tool-use tasks, scoring 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B and leading on instruction-following benchmarks with a 96.75 score on IFEval. The model generates explicit chains of thought before answering questions and integrates with agent frameworks to call external tools within multi-step workflows. Inference speeds are notably fast: the model decodes at 184 tokens per second on Apple's M5 Max and exceeds 21,000 output tokens per second on a single NVIDIA B300 GPU. LiquidAI trained the model on 33 trillion tokens across multiple stages, with specialized post-training phases designed to optimize agentic and tool-use capabilities. The release reflects the industry's shift toward efficient on-device AI. By running entirely on local hardware, the model keeps user data private while eliminating latency from cloud API calls—a growing concern as organizations seek to deploy AI agents at scale. The model ships with support across major inference frameworks including llama.cpp, MLX, vLLM, and SGLang.