Danube-2: The Tiny AI Model Leading the Open LLM Leaderboard

2 May 2024

The AI industry is dominated by discussions of massive LLMs like GPT-4, Claude 3, and Gemini. But there are equally big issues with relying on these closed-source LLMs. For one, you have minimal control. The models often worsen over time and exhibit biases that you can’t fix.

Plus, the firms gatekeeping the models, from OpenAI to Anthropic to Google, can throttle your usage, increase your prices, or simply ban your account. If you run a business, you’re risking everything by using closed-source models. Another issue is their sheer size. They’re far too big and slow to be useful for hardware devices like portable assistants - something Rabbit and Humane AI learned only after scathing reviews.

That’s where open-source, tiny LLM models come into play. The most accurate of these in the <2B parameter category is H2O.ai’s Danube 2. They’re becoming increasingly popular for deployment on edge devices, as traditional large language models are too computationally demanding to run efficiently on resource-constrained hardware.

Tiny LLMs for phones

Danube 2 is a “tiny LLM” that packs a surprising punch for its size. With just 1.8 billion parameters, Danube is lightweight compared to giants like GPT-4, but it achieves highly competitive performance across a wide range of natural language tasks. Trained on a carefully curated dataset of 2 trillion tokens from diverse web sources, Danube leverages techniques refined from models like Mistral to maximize its efficiency.

Basically, you can download this model to your device and run it wherever you like — whether that’s new portable consumer hardware, a security camera, an IoT sensor, or anywhere else.

One of the key advantages of Danube is its accessibility. With an Apache 2.0 license, it’s freely available for commercial use, lowering the barriers to entry for developers and businesses looking to incorporate AI into their mobile applications.

H2O.ai has also released a fine-tuned version, Danube2–1.8B-Chat, specifically optimized for conversational AI. This could spur a wave of more natural and responsive chatbots and virtual assistants on devices from smartphones to smart home hubs.

Tiny AI in healthcare

The potential applications for small AI extend far beyond consumer-facing chatbots or phones. In the healthcare sector, companies like HeartSciences are leveraging compact machine learning models to supercharge the electrocardiogram (ECG).

Heart disease is the leading cause of death globally, claiming around 51,000 lives each day, and ECGs are one of the primary tools for diagnosing cardiac issues. However, interpreting these readings requires specialized expertise that many healthcare providers lack, particularly in resource-constrained settings.

HeartSciences’ MyoVista Wavelet ECG system aims to bridge this gap by incorporating sophisticated AI algorithms capable of spotting subtle patterns that human eyes might miss. Trained on vast datasets, these models act as an early warning system for cardiac dysfunction. Crucially, the machine learning behind MyoVista is designed to run efficiently on the kind of hardware you’d find in a typical doctor’s office or clinic, not a high-performance computing cluster.

This is the essence of small AI — bringing the power of cutting-edge algorithms to the point of care, where they can have the greatest impact on patient outcomes. When minutes can mean the difference between life and death, having AI-powered insights available immediately is a game-changer. And with hundreds of millions of ECGs administered every year, the scalability of lightweight models like those used in MyoVista is key to making a real dent in the global burden of heart disease.

Tiny AI for security

Another promising application for small AI is in the realm of smart security systems. Traditional security cameras rely on motion detection and human monitoring to identify potential threats, leading to high rates of false alarms and missed incidents. However, by incorporating compact machine learning models, these systems can become much more intelligent and autonomous.

Imagine a security camera equipped with an AI model trained to recognize not just motion but specific types of behaviors and objects. Such a system could distinguish between a person walking their dog and someone attempting to break into a building, alerting human operators only when there’s a genuine threat. By running these models directly on the camera hardware rather than streaming video to the cloud for analysis, smart security systems can operate more efficiently and protect sensitive data from potential breaches.

Of course, developing effective AI models for smart security presents its own set of challenges. The models must be able to operate in real-time, processing video feeds as they come in while also being robust to variations in lighting, weather, and other environmental factors. They must also be trained on diverse datasets to avoid bias and ensure fair performance across different demographic groups.

Despite these challenges, the potential benefits of AI-powered smart security are significant. By reducing false alarms and enabling proactive threat detection, these systems can help keep people and property safe while also reducing the workload on human operators. And as the underlying hardware continues to become more powerful and energy-efficient, the possibilities for edge-based AI in security applications will only continue to grow.


As the AI community continues to push the boundaries of what’s possible with large-scale models and infrastructure, it’s important not to lose sight of the transformative potential of small AI. By bringing powerful capabilities to the devices and contexts where they’re needed most, compact models have the potential to democratize access to AI’s benefits and touch billions of lives in meaningful ways.

From the smartphones in our pockets to the medical devices in our clinics to the security cameras watching over our streets, small AI is poised to make a big impact. As we work to overcome the challenges around data, privacy, and bias, we must also recognize and celebrate the quiet revolution that’s happening at the edge. The future of AI may be small, but its potential is enormous.