Run fast, private agents on-device with LFM2.5-24B-A2B
LFM2.5-24B-A2B is a 24B-parameter reasoning model with just 2B active parameters, built for fast, private agents on-device. It thinks before it answers, calls tools, and works through multi-step tasks inside popular agent harnesses.
The Hugging Face Blog announces LFM2.5-24B-A2B, a 24B-parameter reasoning model optimized for fast, private on-device agents with 2B active parameters per token.
What happened
Hugging Face released LFM2.5-24B-A2B, a reasoning model designed for on-device agent applications. It demonstrates competitive performance against larger models while maintaining low computational costs through parameter efficiency.
Why it matters
This model enables developers to deploy complex reasoning agents on edge devices without compromising performance, addressing privacy and latency concerns in AI applications.
- Parameter efficiency enables high performance on edge devices
- Model architecture balances complexity with computational constraints
- Optimized for agent workflows with tool calling capabilities
- Competes with larger models in key benchmarks
- Supports diverse deployment options across hardware platforms
The bigger picture
Reduces reliance on cloud infrastructure for agent-based AI by enabling efficient on-device execution, while maintaining competitive performance in tool use and instruction following tasks.
LFM2.5-24B-A2B offers a compact, efficient solution for on-device AI agents, demonstrating that parameter efficiency can match the performance of larger models in critical tasks.
Technical details
- Mixture-of-Experts (MoE) architecture with 64 experts per block
- 131K token context window with in-place tokenizer expansion
- Hybrid gated short convolutions and GQA (Grouped Query Attention)
- Five-stage post-training process including agentic reinforcement learning
- Support for multiple inference frameworks (llama.cpp, MLX, vLLM, etc.)
FAQ
What makes LFM2.5-24B-A2B unique compared to other models?
It combines 24B parameters with only 2B active parameters per token, enabling efficient on-device reasoning while maintaining competitive performance in agent workflows.
How does the model perform against larger models?
It outperforms Qwen3.6-35B-A3B (50% more parameters) on several benchmarks, including AutomationBench and LiveClaw, while using less memory.
What hardware optimizations does it support?
It achieves 184 tokens/s on Apple M5 Max and 21K output tokens/s on NVIDIA B300 GPU, with support for CPU and GPU inference frameworks.
What are the main use cases for this model?
It's designed for private, on-device agents that require multi-step reasoning, tool calling, and continuous operation without cloud dependency.
How is the model trained?
It uses a five-stage post-training process including supervised fine-tuning, multi-domain distillation, and agentic reinforcement learning across various agent frameworks.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by Hugging Face Blog on Oct 13, 2026.
Comments
Sign in to join the discussion.