Liquid AI Releases LFM2.5-24B-A2B: A 2B Active-Parameter MoE Built for On-Device Agents

On-Device Agents Get a New Option

Liquid AI recently released an open-weights model called LFM2.5-24B-A2B, built on a Mixture-of-Experts (MoE) architecture with 24B total parameters but only 2B activated per inference. This design allows it to run efficiently on local devices, targeting on-device agent scenarios.

According to MarkTechPost, the model scored 48.09 on the LiveClaw benchmark and reaches a decoding speed of 184 tokens/s on an Apple M5 Max. For developers who need to deploy agents locally, these numbers offer an option worth paying attention to.

Why 2B Active Parameters Matter

Traditional dense models need to load all their parameters during inference, placing high demands on memory and compute. MoE architectures use a routing mechanism to activate only a subset of experts at a time, substantially reducing actual computation while preserving a larger total parameter count.

By compressing activated parameters to 2B, LFM2.5-24B-A2B means its inference cost is close to that of a 2B dense model, while its latent knowledge capacity comes from a 24B parameter scale. This "large capacity, small activation" approach is a key direction in current on-device model optimization.

For local agents, response speed and memory footprint often matter more than absolute capability. A decoding speed of 184 tokens/s is already enough to support real-time interactive tasks, such as local Q&A, tool calling, and lightweight automation workflows.

Mixture-of-Experts (MoE) architecture works by splitting the feed-forward network into multiple parallel "expert" sub-networks, then using a lightweight router to dynamically select a small number of experts for each token to participate in computation. Taking LFM2.5-24B-A2B as an example, the 24B total parameter count represents the full weight scale required for storage and loading, while only about 2B parameters (the "activated parameters") actually participate in each forward pass. This means VRAM/memory usage depends on the total parameter count, while per-token computation and latency are closer to the activated parameter count. Compared with the linear relationship of dense models—"more parameters, heavier computation"—MoE gives models much more room to decouple capability from cost, which is the fundamental reason it is widely used in on-device and low-cost inference scenarios.

The Practical Significance of Open Weights and On-Device Deployment

Liquid AI's choice to release open weights lowers the barrier for developers to try and integrate the model. Compared with pure cloud API solutions, open-weights models can run in offline environments, privacy-sensitive scenarios, or on cost-constrained devices.

It is worth noting that the LiveClaw score of 48.09 needs to be understood in the context of the benchmark's task type. It reflects the model's performance on a specific agent evaluation set and does not directly equate to general reasoning or coding ability. When selecting a model, developers should still test it against their own actual workloads.

Open Weights means the parameter files after model training are publicly released, allowing anyone to download them, run them locally, and fine-tune or redistribute them under the license terms—but it is not entirely equivalent to "open source": training data, training code, and the full recipe are often not released together. For on-device deployment, the value of open weights lies in keeping data on the device, enabling fully offline inference, and avoiding per-token cloud fees. Developers can also perform lightweight fine-tuning or quantization (such as 4-bit or 8-bit) on domain-specific data to further compress the model to a size that phones and laptops can handle. Compared with closed-source solutions that only provide an API, it gives up some of the vendor-side operational convenience in exchange for controllability, privacy, and long-term cost autonomy.

The Competitive Landscape for On-Device Agents

The on-device agent track is heating up quickly. Multiple vendors are trying to use MoE, quantization, and miniaturization techniques to fit capabilities that once required cloud compute into laptops and phones.

LFM2.5-24B-A2B's positioning is clear: it is not chasing first place on leaderboards, but rather finding a balance between "being able to run" and "being good enough." For teams that need local deployment, are latency-sensitive, and still want to retain a certain level of reasoning ability, this type of model offers an alternative beyond cloud solutions.

As hardware like Apple Silicon continues to improve memory bandwidth and NPU performance, the practical boundaries of on-device models will keep expanding. LFM2.5-24B-A2B is a concrete example of this trend, and its real-world performance remains to be validated by more developers.

On-Device Agent refers to an AI system that runs locally on a phone, laptop, or edge device and can autonomously plan and call tools to complete multi-step tasks, rather than only handling single-turn Q&A. Its requirements for models differ noticeably from cloud agents: it must not only understand instructions, but also complete tool calls, state tracking, and multi-turn decision-making under memory constraints and without stable network access, so low latency and stable structured output are often more critical than absolute intelligence level. Model quantization, MoE sparse activation, and unified memory architectures like Apple Silicon are several key technologies supporting this scenario. When evaluating such models, agent-specific benchmarks (such as LiveClaw mentioned in this article) better reflect real usability than general Q&A leaderboards.

One-line Summary

Liquid AI releases a 24B-total, 2B-active MoE open-weights model focused on on-device agents.

Paragraph Summary

Liquid AI released the open-weights model LFM2.5-24B-A2B, using a Mixture-of-Experts architecture with 24B total parameters and only 2B activated per inference, targeting on-device agent scenarios. The model scored 48.09 on the LiveClaw benchmark and reaches a decoding speed of 184 tokens/s on an Apple M5 Max. The article argues that this "large capacity, small activation" design brings inference cost close to a 2B dense model while retaining the knowledge capacity of a 24B scale, matching local agents' need to prioritize response speed and memory footprint over absolute capability. The open-weights release lowers integration barriers and supports offline and privacy-sensitive scenarios. However, the 48.09 score must be understood in the context of the benchmark's task type and is not equivalent to general reasoning ability; developers should still test against actual workloads. The on-device agent track is heating up quickly, and this model is positioned between "being able to run" and "being good enough."