LiquidAI launches LFM2.5-24B-A2B AI agent model

LiquidAI launches LFM2.5-24B-A2B, an AI agent model optimized for on-device use. The model is compared with Qwen3.6-35B-A3B, with competitive performance in agentic tasks and tool use.

It is optimized for applications that require tool use and instruction following. The model has 24 billion parameters and is capable of reasoning at MoE speed, with 2B parameters per token. The LFM2.5-24B-A2B scored 48.09 on LiveClaw, versus 38.19 for Qwen. It offers inference efficiency, with 184 tok/s on the M5 Max and up to 21K tok/s on the B300. The model's efficiency allows it to run on devices with limited resources without compromising performance.

Availability and pricing in Brazil were not disclosed. The model is aimed at developers and users who need an efficient AI agent on devices. The information was revealed by huggingface. More parameters mean greater flexibility in handling complex tasks, such as logical reasoning and decision-making. This allows it to answer queries more accurately and quickly, even in situations that require multiple steps of analysis. This difference indicates more robust performance in tasks involving agentic behavior and tool use.

Inference efficiency is also notable, with up to 21K tok/s on the B300, which makes the model faster at data processing than many in its category. These numbers show that the LFM2.5-24B-A2B is a solid option.

Specifications

Specifications
Item	Specification
Performance on LiveClaw	48.09
Qwen's performance on LiveClaw	38.19
MoE-speed reasoning	2B active parameters per token
Training tokens	33T

Comparison with a rival model and the evolution of the lineup

The competition among AI agent models is fierce, and LiquidAI is positioning itself with a proposal that combines performance and on-device optimization. The LFM2.5-24B-A2B model is compatible with software such as vLLM, llama.cpp, MLX, SGLang and ONNX, which broadens its possible uses. However, to run on devices, the model depends on compatible software, which may limit its use in less technical settings. Availability and pricing in Brazil have not been confirmed, but the comparison with the Qwen3.6-35B-A3B suggests that LiquidAI is aligning with a trend toward more efficient and versatile AI agent models.

The LFM2.5-24B-A2B prioritizes processing efficiency over parameter density, allowing devices to run complex AI models with a lighter load. This matters for users who need an efficient AI agent on laptops and tablets.