Skip to main content

Liquid AI's New Model Brings Powerful AI to Your Phone

AI Gets Personal: Liquid AI Releases Powerful Edge-Side Model

Artificial intelligence just became more personal. Startup Liquid AI has open-sourced its latest breakthrough - the LFM2.5-8B-A1B model that brings desktop-quality AI to everyday devices. This isn't just another incremental update; it's a significant leap forward in making powerful AI accessible anywhere.

Image

Smart Design for Smart Devices

The magic lies in its architecture. Using a sparse mixture-of-experts (MoE) approach, the model packs 8.3 billion parameters but only activates 1.5 billion per task. Think of it like having a team of specialists where only the relevant experts weigh in on each question - making it remarkably efficient without sacrificing capability.

"We've essentially created a model that thinks before it speaks," explains the development team. Before delivering final answers, LFM2.5 generates explicit reasoning chains, helping avoid those frustrating AI hallucinations we've all encountered.

More Context, Fewer Mistakes

Compared to its predecessor, the improvements are substantial:

  • Context window quadrupled to 128K tokens
  • Training data more than tripled to 38T tokens
  • Enhanced multilingual support covering nine languages

But raw numbers don't tell the whole story. The team implemented innovative two-stage reinforcement learning to tackle persistent AI challenges. One system optimizes for logical consistency in long reasoning chains, while another actively prevents the model from answering beyond its knowledge - a feature many users will appreciate.

Performance That Surprises

Benchmark tests reveal performance that rivals larger models, particularly in logical reasoning and instruction following. What makes this impressive is that it achieves these results while being designed to run on your phone or laptop.

Tool calling gets special attention too. The model defaults to efficient Python function calls but can seamlessly switch to JSON format when needed - flexibility that developers will love.

Ready When You Are

Unlike some AI releases that leave users waiting for ecosystem support, LFM2.5 launched with immediate compatibility across major platforms including llama.cpp, MLX, vLLM, and SGLang. Early tests show it decoding at 253 bytes per second on M5 Max chips, with mobile devices hitting about 30 bytes per second - more than enough for responsive local AI applications.

Key Points:

  • Liquid AI's LFM2.5 brings advanced AI to consumer devices
  • Sparse expert architecture delivers efficiency without compromising power
  • 128K token context and anti-hallucination features set new standards
  • Full ecosystem support available at launch
  • Performance benchmarks rival larger models while running locally