AI Tools & Products
Oct 14, 2026
Launch of LFM2.5-24B-A2B Model for Efficient On-Device AI Agents
Oct 14, 2026
AI Summary
The LFM2.5-24B-A2B model has been released, featuring 24 billion parameters with 2 billion active for enhanced on-device performance. It competes effectively with larger models in various benchmarks, demonstrating significant speed and efficiency for AI applications.
-oOdUn61bDRSykk4H951SoDOHzqoOe1.png)
- LFM2.5-24B-A2B is a new AI model with 24 billion parameters, designed for fast on-device use with only 2 billion active parameters.
- It is the largest model in the LFM2.5 series, which includes models ranging from 230 million to 24 billion parameters.
- The model shows competitive performance against larger models, such as Qwen3.6-35B-A3B, on several benchmarks, scoring 48.09 on LiveClaw and outperforming on AutomationBench.
- LFM2.5-24B-A2B runs up to 1.7 times faster than similar models, achieving 184 tokens per second on an Apple M5 Max and over 21,000 tokens per second on an NVIDIA B300 GPU.
- It features an expanded vocabulary and context window, with improvements in pre-training and reinforcement learning strategies.
- The model is open-weight and available on platforms like Hugging Face, with documentation for local running and fine-tuning.
- It is designed for tasks requiring reasoning, tool usage, and reliable performance on personal devices, with recommendations for pairing with retrieval systems for knowledge-heavy tasks.
- The architecture maintains 40 layers and 64 experts per MoE block, with enhancements to the context window and tokenizer.
- LFM2.5-24B-A2B has been evaluated across 16 benchmarks, showing strong performance in tool use, instruction following, and agentic workflows.
- The model is positioned as a foundation for AI agents that operate efficiently on local hardware, with a focus on interactive task management without cloud dependency.
on-devicefast agentsliquid aitechnologyinnovation