Skip to main content

Liquid AI's New Model Runs Entirely on Your Phone, No Cloud Needed

In a move that could reshape how we think about AI on our devices, Liquid AI has just dropped LFM2.5-2.6B—a compact yet powerful model designed to run entirely on your smartphone. No cloud, no API calls, just pure on-device intelligence. Announced on August 5, this 2.6-billion-parameter model is tailored for agent workflows, making it a game-changer for developers who want speed, privacy, and cost efficiency.

What Makes It Special?

At its core, LFM2.5-2.6B is all about bringing AI closer to where you are. By running locally, it sidesteps the latency and privacy concerns that come with sending data to the cloud. Imagine asking your phone to book a restaurant, manage your calendar, or even draft an email—all without your personal data ever leaving your device. That's the promise here.

But how does it stack up technically? The model was pre-trained on a whopping 34 trillion tokens, with a vocabulary expanded to 128K. That's a lot of data, and it shows. The post-training process is equally impressive, involving four stages: supervised fine-tuning, teacher specialization, multi-domain strategy distillation, and agent reinforcement learning. The result? A model that excels at planning, tool calling, and handling complex, multi-step tasks.

Performance That Surprises

You might expect a smaller model to lag behind its larger counterparts, but LFM2.5-2.6B defies expectations. In benchmark tests, it matches the performance of models nearly four times its size in instruction following and tool use. That's no small feat.

On the hardware front, it's equally impressive. On Apple's M5 Max chip, it hits a blazing 220 tokens per second. Even on mobile devices, it maintains a steady 30 tokens per second—plenty for real-time interactions. So whether you're on a high-end laptop or a mid-range phone, you're getting a responsive, snappy experience.

Open Source and Ready to Go

One of the most exciting aspects is that LFM2.5-2.6B is fully open-sourced on Hugging Face, along with its base version. That means developers can dive in, customize, and deploy it right away. It natively supports popular inference frameworks like llama.cpp, MLX, and vLLM, so integration is a breeze.

For developers, this is a golden opportunity. You can build edge-side agents that are not only efficient but also respect user privacy. And with the cost savings from not relying on cloud infrastructure, it's a win-win.

The Bigger Picture

Liquid AI's move signals a broader trend toward edge computing in AI. As models get smaller and smarter, the need for constant cloud connectivity diminishes. This could lead to faster, more private, and more accessible AI applications across the board.

But it's not just about the tech. It's about what this means for everyday users. Imagine a world where your devices understand you better, respond faster, and keep your data safe. That's the future LFM2.5-2.6B is helping to build.

Key Points

  • Liquid AI has released LFM2.5-2.6B, a 2.6-billion-parameter model for edge devices.
  • It runs entirely locally, eliminating the need for cloud APIs.
  • Performance is comparable to models 4x its size in instruction following and tool use.
  • Achieves 220 tokens/second on M5 Max, 30 tokens/second on mobile.
  • Open-sourced on Hugging Face, supporting llama.cpp, MLX, and vLLM.