LFM2.5-24B-A2B: Built for fast on-device agents (www.liquid.ai)

Today marks the release of LFM2.5-24B-A2B, a state-of-the-art 24 billion parameter reasoning model designed specifically for fast, on-device performance, utilizing only 2 billion active parameters. This model, which is the largest in the LFM2.5 series, demonstrates impressive capabilities, outperforming larger models like Qwen3.6-35B-A3B on several benchmarks related to agentic tasks and tool usage. With a remarkable decoding speed of 184 tokens per second on an Apple M5 Max and a staggering 21,000 output tokens per second on an NVIDIA B300 GPU, LFM2.5-24B-A2B is optimized for real-time applications, emphasizing privacy by enabling local processing without cloud dependencies. The advancements in LFM2.5-24B-A2B build upon its predecessor with an expanded context window from 32,768 to 131,072 tokens and a vocabulary increase from 65,000 to 128,000 tokens. It employs a novel reasoning approach that produces explicit chains of thought, enhancing answer quality. The model's training included a large-scale reinforcement learning phase and was benchmarked against various workloads, demonstrating its strengths in instruction-following and agentic scenarios. LFM2.5-24B-A2B is available on platforms like Hugging Face and comes with tools for local deployment, making it an appealing choice for developers seeking efficient, intelligent agents that operate seamlessly on consumer hardware.