LiquidAI releases LFM2.5-24B-A2B reasoning model for on-device agents
LFM2.5-24B-A2B is a 24B-parameter Mixture-of-Experts model that uses only 2B active parameters to enable fast, private, on-device reasoning and tool use.
Summary
LiquidAI has released LFM2.5-24B-A2B, an open-weight reasoning model designed for efficient on-device agentic workflows. Despite having 24B total parameters, the Mixture-of-Experts (MoE) architecture activates only 2B parameters per token, significantly reducing computational costs. The model features an expanded 131K token context window and demonstrates competitive performance in tool use and instruction following compared to much larger models.
Technical specifications of LFM2.5-24B-A2B
- Architecture: 40 layers with 64 experts per MoE block using top-4 routing, combining gated short convolutions and GQA.
- Efficiency: Achieves 184 tok/s on Apple M5 Max and up to 21K tok/s on a single NVIDIA B300 GPU.
- Context & Vocabulary: Context window increased to 131K tokens; vocabulary expanded from 65K to 128K tokens.
- Training: Pre-trained on ~33T tokens, with additional mid-training on reasoning, math, and code.
- Compatibility: Supports llama.cpp, MLX, vLLM, SGLang, and ONNX.
Why it matters
The model provides a high-performance, low-resource option for developers building private, local AI agents using open-source frameworks.
Original reference: huggingface.co →