Skip to main content

Liquid AI's New Model Runs Locally, No Cloud Needed

In an era where AI often feels tethered to the cloud, Liquid AI is charting a different course. On August 5, the startup announced LFM2.5-2.6B, a compact yet powerful model built specifically for on-device agent workflows. With 2.6 billion parameters, it runs locally on smartphones and other terminal devices, bypassing the need for cloud APIs altogether. For developers, that means lower costs, reduced latency, and a stronger grip on user privacy.

What makes this model tick? Liquid AI trained LFM2.5-2.6B on a staggering 34 trillion tokens, expanding its vocabulary to 128K. But the real magic lies in its post-training regimen, which blends four stages: supervised fine-tuning, teacher specialization, multi-domain strategy distillation, and agent reinforcement learning. This cocktail equips the model with sharp planning abilities, reliable tool calling, and the knack for juggling complex, multi-step tasks. In benchmark tests, it holds its own against models nearly four times its size—a testament to its efficiency.

Hardware compatibility is another feather in its cap. On Apple's M5Max chip, it churns out 220 tokens per second, while mobile devices still manage a respectable 30 tokens per second. That's not just fast; it's practical for real-world applications where responsiveness matters.

Liquid AI has also embraced openness. Both LFM2.5-2.6B and its base version are now available on Hugging Face, with native support for popular inference frameworks like llama.cpp, MLX, and vLLM. This move lowers the barrier for developers eager to experiment with edge AI.

So, what does this mean for the broader AI landscape? It signals a shift toward more autonomous, privacy-conscious applications. Imagine a voice assistant that never sends your queries to a server, or a health app that analyzes data on your phone without uploading it. These scenarios are inching closer to reality.

Of course, challenges remain. Edge models often sacrifice some accuracy for speed and efficiency, and LFM2.5-2.6B is no exception. But for many use cases, the trade-off is worth it. As Liquid AI continues to refine its approach, we might see a future where AI is not just smart, but also discreet and responsive—right in the palm of your hand.

Key Points

  • Local Execution: LFM2.5-2.6B runs entirely on-device, eliminating cloud API dependencies.
  • Performance: Outperforms models four times its size in instruction following and tool use.
  • Speed: Achieves 220 tokens/sec on M5Max, 30 tokens/sec on mobile.
  • Open Source: Fully available on Hugging Face, supporting llama.cpp, MLX, and vLLM.
  • Privacy & Cost: Offers low latency and enhanced privacy for edge applications.