Liquid AI's LFM2.5-230M: Tiny Model, Big Performance for On-Device AI

Liquid AI Brings Powerful On-Device AI to Your Pocket

Liquid AI just released LFM2.5-230M, a remarkably compact open-weight model that's reshaping what's possible with on-device artificial intelligence. At just 230 million parameters, this model proves you don't need billions of parameters to get useful, fast AI—and that's transformative for anyone working with edge devices.

What Makes This Release Significant?

The headline numbers tell part of the story: the model runs at 213 tokens per second on a Galaxy S25 Ultra and 42 tokens per second on a Raspberry Pi 5. These aren't theoretical benchmarks; they're real-world performance on devices most people actually own or can afford. For context, that's the difference between a chatbot feeling instant versus watching paint dry.

What's equally important is what it can do. Built on the proven LFM2 architecture, LFM2.5-230M specializes in tool use and data extraction—two capabilities critical for practical AI applications. In benchmark tests, it outperforms notably larger models like Qwen3.5-0.8B and Gemma 3 1B on instruction following tasks.

Why This Matters for the Broader AI Landscape

We're witnessing a quiet shift in AI development philosophy. For years, the narrative centered on bigger is better—larger models, more parameters, more compute. LFM2.5-230M challenges that assumption. By proving that a 230M model can beat 800M and 1B models at specific tasks, Liquid AI demonstrates that efficiency and architecture matter as much as raw scale.

This has real implications:

• Privacy and control: Running models directly on personal devices means data never leaves your phone or local server
• Cost reduction: No cloud API calls mean lower expenses for developers and businesses
• Latency elimination: Instant responses without network round-trips
• Accessibility: Powerful AI becomes viable on budget hardware and aging devices

Framework Support That Matters

The release includes support for llama.cpp, MLX, vLLM, SGLang, and ONNX—a comprehensive toolkit that makes deployment straightforward. Whether you're a developer building on Apple Silicon (MLX), optimizing for CPUs (llama.cpp), or managing inference servers (vLLM), there's a path forward. This broad compatibility signals that Liquid AI understands real-world deployment challenges.

What This Means for AI Tool Users

If you're evaluating AI tools or platforms, this release expands your options significantly. Small, efficient models open doors that massive ones couldn't:

• Mobile app developers can now integrate capable AI without ballooning app sizes
• Edge computing companies gain a proven model for resource-constrained environments
• Privacy-conscious organizations can deploy advanced AI without external dependencies
• Cost-sensitive operations reduce infrastructure spending while maintaining capability

For everyday users, this cascades into better mobile experiences, faster local AI assistants, and smarter edge devices—all without sacrificing capability.

The Bigger Picture

LFM2.5-230M is one data point, but it represents an encouraging trend: the AI community is learning to build smart, not just big. As models become more efficient and accessible, AI tools become ubiquitous rather than centralized—a genuinely different vision for how AI shapes our future.

The Takeaway

Liquid AI's latest release proves that competitive on-device AI doesn't require massive models or expensive infrastructure. For tool builders, platform developers, and anyone concerned about privacy, cost, or latency, LFM2.5-230M represents a meaningful step forward. This is the kind of efficiency innovation that actually changes what's possible in practice.