Insider Brief
- Multiverse Computing has introduced the “Model Zoo,” a family of ultra-compact AI models that run locally while matching or exceeding the performance of larger systems.
- The ChickenBrain model, a compressed version of Llama 3.1, is 3,700 times smaller than typical requirements and outperforms the original on multiple industry benchmarks.
- The SuperFly model, based on SmolLM2 135, has 94 million parameters and enables offline conversational AI for uses such as smart appliances and in-vehicle assistants.
Multiverse Computing says it has cracked a problem that has long dogged artificial intelligence: making models smaller without losing performance, according to a company news release.
The Spain-based company on Thursday unveiled what it calls the “Model Zoo,” a lineup of “nano models” that it claims can run locally on modest hardware while matching or exceeding the performance of much larger systems. The release includes two new products — ChickenBrain and SuperFly — designed to show that bigger isn’t always better in AI.
Small Brains, Big Results
ChickenBrain is a stripped-down version of Meta’s Llama 3.1 large language model that Multiverse says is 3,700 times smaller than the computing resources typically required. Despite the cut, the model outperformed Llama 3.1 8B on industry benchmarks such as MMLU Pro, MATH500, GSM8K, and GPQA Diamond. The company tested it on everyday devices like a MacBook Pro and a low-cost Raspberry Pi, showing it could run without specialized cloud infrastructure.
SuperFly, even smaller, is based on the open-source SmolLM2 135 model and has just 94 million parameters. Multiverse says it is tiny enough to fit within the neural capacity of two flies, or 15,000 times smaller than a chicken brain. The model is intended for conversational AI that can run without an internet connection, making it suitable for embedding into consumer products and vehicles.
The company points to immediate uses in areas such as smart home appliances, where SuperFly could enable voice controls without sending data to the cloud. That would allow a washing machine or refrigerator to respond to plain-language commands, even offline. In cars, it could power onboard assistants that continue to operate in areas without cellular coverage, handling navigation, climate, and audio tasks.
Breaking From the Pack
Much of the AI industry has focused on reducing model size using methods like quantization, pruning, and knowledge distillation. Multiverse says its quantum-inspired compression techniques go further, achieving results that defy the traditional trade-off between size and performance.
The company argues that the push for ever-larger models has created an energy and cost problem that smaller, more efficient systems could solve. “The future of AI,” it says, “does not lie in ever-larger models and increasingly demanding hardware.”
ChickenBrain and SuperFly are available by private request. The broader Model Zoo will be accessible through Multiverse’s CompactifAI API in the coming months.
