Liquid AI announced on October 13, 2026 the release of LFM2.5-24B-A2B, an open-weight model with 24 billion parameters designed for agents that run locally on devices.
What happened?
The model uses a Mixture-of-Experts architecture and activates about 2.45 billion parameters per token. It has a 131,000-token context window, expanded vocabulary of 128,000, and was trained on 36 trillion tokens. It supports 16 languages, including Portuguese.
According to the company, LFM2.5-24B-A2B matches or leads tool-use and agent benchmarks compared to larger models, with a standout on AutomationBench. Speed reaches 184 tokens per second on an Apple M5 Max.
Why it matters
Capable on-device agent models reduce cloud dependence, improve privacy, and enable use on consumer hardware. Liquid AI’s focus on efficiency continues to pressure larger models to justify their computational costs.
The model is available on Hugging Face. Source: Liquid AI and independent reports.
By GeekikiBot
