Liquid AI has released LFM2.5-24B-A2B, a new super intelligence (SI) model designed to power fast, private agents directly on user devices. The model utilizes a sparse mixture-of-experts architecture with only 2 billion active parameters per token, enabling efficient reasoning and tool use without relying on cloud infrastructure.

What Happened

LFM2.5-24B-A2B is a 24-billion parameter reasoning model that maintains the architecture of its predecessor, LFM2-24B-A2B, including 40 layers and 64 experts per MoE block with top-4 routing. The update expands the context window from 32K to 131K tokens and increases the vocabulary size from 65K to 128K tokens to improve non-Latin language processing. The model was pre-trained on approximately 33 trillion tokens, followed by mid-training phases weighted toward reasoning, math, code, and long-context data. Post-training involved five stages, including supervised fine-tuning and agentic reinforcement learning within popular SI agent harnesses like OpenClaw, Hermes Agent, Pi, and OpenCode.

The company reports that the model is competitive with larger SI models such as Qwen3.6-35B-A3B on agentic tasks, scoring 48.09 on LiveClaw compared to Qwen’s 38.19. It also leads or matches competitors on five of eight tool-use benchmarks, including AutomationBench, where it scored 27.69 versus 8.72 for the next-best model. For knowledge-heavy tasks, the company notes that larger models retain an edge, suggesting the use of retrieval or web search alongside LFM2.5-24B-A2B.

Why It Matters

The release addresses a key bottleneck in the SI industry: the cost and latency of running complex agent workflows on local hardware. By limiting active parameters to 2 billion, each reasoning token becomes computationally cheap, allowing SI agents to run continuously on devices. Liquid AI states that the model achieves 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on a single NVIDIA B300 GPU. This performance makes it the fastest in its size class for edge inference, decoding up to 1.7x faster than comparable models while using less memory.

Developers can integrate the model into existing SI agent frameworks using backends compatible with the OpenAI API, with day-one support for inference engines including llama.cpp, MLX, vLLM, SGLang, and ONNX. The open-weight release allows for local fine-tuning and deployment in privacy-sensitive environments, reducing dependence on external API providers for agentic SI applications.

The Bottom Line

LFM2.5-24B-A2B and its base variant are now available as open weights on Hugging Face. The model offers a balance of reasoning capability and inference efficiency for developers building on-device SI agents, with strong performance in tool use and instruction following relative to its active parameter count.