Open models, honest benchmarks, and demos you can try on real data.
Three open models under 1B parameters vote on an intent and return a confidence you can act on. Zero-shot, no training, runs on a CPU.