LiquidAI has announced the release of LFM2.5-24B-A2B, the largest model in its LFM2.5 family. The new model is a 24-billion-parameter reasoning model designed specifically for on-device deployment, utilizing a Mixture-of-Experts (MoE) architecture that maintains only 2.45 billion active parameters per token.
Technical Specifications and Architecture
The LFM2.5-24B-A2B is a general-purpose, text-only model trained on a budget of 36 trillion tokens. According to technical documentation provided by the company, the architecture consists of 40 layers, comprising 30 double-gated short convolution blocks and 10 Grouped Query Attention (GQA) layers. The model utilizes 64 experts per MoE block with top-4 routing.
The model supports a context length of 131,072 tokens and features a vocabulary size of 128,000. It is designed for multilingual applications, supporting 16 languages including English, Chinese, Arabic, French, German, Japanese, and several others.
Performance Benchmarks and Efficiency
LiquidAI claims the model is optimized for “fast, private agents on-device.” In comparative testing, the company reports that LFM2.5-24B-A2B performs competitively with the Qwen3.6-35B-A3B model on agentic and tool-use tasks. Specifically, the company cites a score of 48.09 on the LiveClaw benchmark, compared to 38.19 for the Qwen model, despite the Qwen model having approximately 50% more total and active parameters.
In terms of inference speed, LiquidAI states the model runs up to 1.7× faster than comparable MoE models on-device. Reported performance metrics include 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on a single NVIDIA CORP B300 GPU.
Availability and Deployment
The model is available via Hugging Face in GGUF format, facilitating local deployment. LiquidAI also provides a free demo through its playground interface.
Discussions on community platforms such as Reddit indicate interest in the model’s potential for agentic workflows, such as file editing, where users have noted limitations in other existing models like Gemma. While some users have expressed interest in speculative decoding via a “DSpark” draft model for faster generation, this feature is currently listed as “coming soon” and has not yet been released.