Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE for On-Device Agents
What it does
Liquid AI released the LFM2.5-24B-A2B, an open-weight Mixture-of-Experts (MoE) reasoning model with 24 billion parameters, but only 2 billion active parameters on each pass. This design uses sparse activation to limit compute requirements while maintaining model size and capacity. The model targets local, on-device AI agents rather than cloud-based systems. It scored 48.09 on the LiveClaw benchmark and processes 184 tokens per second on an Apple M5 Max chip.
Why it matters
MoE models usually require large server infrastructure due to their massive parameter counts and compute needs. By activating only 2 billion parameters out of 24 billion, Liquid AI’s model dramatically reduces the real-time resource demands, making sophisticated reasoning feasible on consumer-grade hardware like laptops or edge devices. This approach shifts AI power closer to users, lowering latency and improving privacy by avoiding cloud dependency.
The LiveClaw score suggests competitive reasoning ability, and the throughput on Apple silicon indicates practical speeds for interactive applications. Builders working on personal AI assistants or on-device analytics now have a viable option that balances model complexity with hardware constraints.
Who it is for
This model primarily serves developers building local AI agents for devices such as laptops, tablets, or embedded systems with constrained compute budgets. It suits startups focusing on privacy-first AI or businesses aiming to reduce cloud costs and latency for real-time AI features. Device manufacturers can also integrate the model into user-facing agents without relying heavily on cloud API calls.
The catch
The LFM2.5-24B-A2B model, while open-weight, likely demands advanced engineering to implement and optimize on non-standard hardware. Performance gains hinge on efficient sparse activation routing, which can be complex to manage. The 2 billion active parameters reduce model computation but do not eliminate the need for powerful chips like Apple’s M5 Max to reach practical decoding speeds. Also, benchmarks like LiveClaw illustrate reasoning but do not capture all real-world use cases, so performance in specific applications may vary.
What to watch next
The impact of LFM2.5-24B-A2B depends on how quickly builders adopt and integrate it into commercial or personal AI agents. Watch for new products leveraging this model to deliver local AI reasoning with minimal cloud support. Also track how Liquid AI or others refine sparse activation techniques on different hardware platforms. Competitive responses from other MoE model developers focused on on-device inference will pressure pricing, accessibility, and performance standards for local agent AI.
AI Quick Briefs Editorial Desk