Multiverse unveils tiny AI models sized after fly and chicken brains

Running a high-performing reasoning AI model offline on your laptop—or embedding a voice assistant inside a washing machine with zero internet connection—is getting closer to reality.

European AI startup Multiverse Computing just rolled out two ultra-compact AI models cleverly grouped under its "Model Zoo" banner. True to the name, they are dubbed after animal brains: SuperFly and ChickBrain.

Multiverse claims these are the world’s smallest high-performing models capable of handling local chat, speech, and even complex reasoning directly on-device.

Quantum-inspired model shrinking

Multiverse didn't build these models from scratch. Instead, it shrank existing open weights using CompactifAI, a compression technology rooted in quantum physics rather than traditional computer science techniques.

SuperFly comes in at an absurdly small 94 million parameters, compressed down from Hugging Face’s 135M SmolLM2. Founder Román Orús described it as basically a fly’s brain, "just a little bit more clever." Designed for ultra-low compute setups like Arduino boards, Multiverse envisions it handling voice commands for appliances—like asking an oven how to fix an error code.

ChickBrain is the heavyweight of the pair at 3.2 billion parameters, squeezed out of Meta’s Llama 3.1 8B. Despite losing over half its size, Multiverse says internal benchmarks show it slightly outperformed the original 8B model across MMLU-Pro, Math 500, GSM8K, and GPQA Diamond.

The Spanish startup, which raised €189 million in June from backers like HP Tech Ventures and Toshiba, is already pitching hardware giants. Orús confirmed ongoing discussions with Apple, Samsung, and Sony. Beyond direct on-device licensing, Multiverse is also serving these slimmed-down models via an AWS API at discounted token rates.

Why it matters

We think Multiverse is attacking the single biggest roadblock in edge AI: the sheer bloat of today's models. Cloud-dependent smart home devices are plagued by latency, server costs, and privacy headaches. If an appliance can handle natural voice commands entirely offline on cheap chips, it changes the economics of consumer hardware.

For developers, ChickBrain looks like a sweet spot. Being able to run reliable reasoning workflows locally on a baseline MacBook without paying per-token API fees or pinging a remote server is a massive win for privacy-sensitive apps.

That said, healthy skepticism is warranted. ChickBrain's benchmark gains over the original Llama 8B are strictly based on internal Multiverse testing, and a 94M model like SuperFly will inevitably be fragile outside narrow, pre-scripted domains. But by proving that quantum-derived compression can shrink models without trashing their cognitive abilities, Multiverse is carving out a genuinely interesting niche against traditional quantizers.