LFM2.5-24B-A2B: Built for fast on-device agents
Source Entity
Hacker News
The LFM2.5-24B-A2B is a new 24B-parameter model with 2B active parameters, optimized for high-speed, private on-device agent tasks. It outperforms larger competitors like Qwen3.6-35B-A3B in key benchmarks while offering significantly faster processing speeds.
The Evolution of On-Device Reasoning: Introducing LFM2.5-24B-A2B
The release of the LFM2.5-24B-A2B marks a significant milestone in the development of efficient artificial intelligence. As the final and largest iteration in the LFM2.5 family, this model bridges the gap between massive, cloud-dependent neural networks and the practical, private requirements of on-device computing. By utilizing a 24B-parameter architecture with only 2B active parameters, the developers have successfully navigated the trade-off between model capacity and computational overhead.
Architectural Efficiency and Sparse Activation
At the heart of the LFM2.5-24B-A2B is its sparse activation structure. By activating only 2B parameters per token, the model achieves a level of agility that dense models of similar size cannot match. This architectural choice is specifically engineered for 'agentic' workflows—tasks that require reasoning, tool-use, and real-time interaction. This efficiency is not merely theoretical; it allows the model to maintain the depth of a 24B-parameter intelligence while operating with the speed of a much smaller system.
Benchmarking Against Industry Standards
In comparative testing, the model demonstrates remarkable performance, particularly when measured against the Qwen3.6-35B-A3B. Despite having approximately 50% fewer total and active parameters, the LFM2.5-24B-A2B shows clear superiority in agentic and tool-use scenarios. Specifically, its score of 48.09 on the LiveClaw benchmark, compared to the competitor's 38.19, highlights its refined capability in handling complex, multi-step digital tasks that are essential for modern AI agents.
Unprecedented Speed on Consumer and Enterprise Hardware
The performance gains are most apparent when examining raw speed metrics. The model achieves 184 tokens/s on Apple’s M5 Max hardware, a feat that makes high-end, private AI agents viable for consumer devices. Furthermore, the capacity to reach 21K output tokens/s on a single NVIDIA B300 GPU underscores its utility in server-side scaling. This speed is a critical factor for developers aiming to build responsive, real-time interfaces without the latency associated with cloud round-trips.
Implications for Privacy and Open-Weight Ecosystems
By releasing both the base and post-trained versions as open-weight models, the team behind LFM2.5-24B-A2B is fostering a more decentralized AI ecosystem. This move empowers developers to integrate sophisticated, private reasoning agents directly into applications, reducing reliance on third-party APIs. As the final entry in the LFM2.5 series—which spans from the lightweight 230M version to this 24B powerhouse—the model represents a mature, scalable solution for the future of on-device intelligence.
Future Trends in Edge AI
The trajectory of the LFM2.5 family suggests that the industry is moving toward a future where 'agentic' intelligence is localized. As these models become more efficient, we can expect a shift in how software interacts with users, moving from static command-response systems to proactive, reasoning-based assistants that operate entirely locally. The LFM2.5-24B-A2B sets a high bar for this transition, proving that massive parameter counts are no longer the sole metric for intelligence.