Brazilian Lab's 144M CPU Decision Model

🔥 What happened Supersonic Labs, a small Brazilian AI lab, dropped Julia 1: a 144.3M-parameter decision model that runs on a plain CPU. You feed it 2–20 candidate answers; it picks one and returns softmax probabilities. No text generation. Apache 2.0 weights on Hugging Face, plus an ONNX/WebGPU build. 💡 Why it matters Training cost just US$104 in cloud GPUs. Median latency on an Apple M4 is 33.15 ms per decision. But on Banking77 with 72 labels it collapses to 64% vs. 87% reference — a clear failure. ⚡ Our take A $104 model that runs on CPU is a warning shot to API pricing. But 72-label routing breaks it — check your use case before ditching the cloud.