Run fast, private agents on-device with LFM2.5-24B-A2B
Source Entity
Hugging Face - Blog
The new LFM2.5-24B-A2B model represents a breakthrough in on-device AI, offering high-performance agentic capabilities with superior efficiency. By utilizing a 2B active parameter architecture, it achieves competitive reasoning speeds and tool-use accuracy on consumer and enterprise hardware.
The Emergence of LFM2.5-24B-A2B
The landscape of artificial intelligence is currently undergoing a pivotal shift toward edge computing, and the release of LFM2.5-24B-A2B marks a significant milestone in this transition. By packing robust agentic reasoning into a 24B-parameter model with only 2B active parameters per token, developers are effectively bridging the gap between high-end cloud performance and the privacy-centric requirements of on-device execution.
Competitive Agentic Performance
One of the most striking aspects of LFM2.5-24B-A2B is its performance on the LiveClaw benchmark. Scoring 48.09 compared to the 38.19 achieved by the Qwen3.6-35B-A3B, the model demonstrates that parameter count is not the sole determinant of intelligence. Despite the Qwen model possessing roughly 50% more total and active parameters, LFM2.5-24B-A2B excels in complex agentic and tool-use tasks, proving that architectural efficiency can outperform brute-force scale.
The Economics of Reasoning
Beyond raw accuracy, the model introduces a 'reasoning at speed' paradigm. By limiting active parameters to 2B per token, the computational cost of the 'thinking' process—where the model evaluates logic before generating an answer—is drastically reduced. This makes complex reasoning tasks, which were once exclusively the domain of massive data centers, economically viable for local deployment.
Hardware Efficiency and Inference
Inference speed is the final hurdle for any on-device model, and LFM2.5-24B-A2B addresses this with impressive hardware optimization. Achieving 184 tokens per second on an Apple M5 Max and reaching up to 21,000 output tokens per second on an NVIDIA B300 GPU, the model is built to scale across various hardware ecosystems. This versatility allows it to serve as a backbone for privacy-focused AI agents that require near-instant response times.
Broader Implications for Privacy and AI Agents
The trend toward on-device agents is a direct response to growing user concerns regarding data privacy. By keeping sensitive reasoning and tool-use operations local, LFM2.5-24B-A2B mitigates the need for external server communication. This shift is likely to accelerate the adoption of autonomous agents in industries such as finance, healthcare, and personal productivity, where data sovereignty is of paramount importance.
Future Trends in Edge AI
Looking forward, the success of models like LFM2.5-24B-A2B suggests that the future of LLMs lies in specialized, dense-active, sparse-model architectures. As developers continue to refine the ratio between active and total parameters, we can expect even smaller devices to handle increasingly complex cognitive tasks. This trajectory points toward a future where the 'brain' of an AI agent is permanently integrated into the local device, ensuring constant, private, and high-speed utility for the end-user.