Jeff: Tiny Models Beat Jev

🔥 What happened Jeff, an independent project, dropped tiny fine-tunes of Qwen3.5 and Gemma 4 (0.8B and 2B) for zero-shot classification. They return calibrated probabilities in a single forward pass — 22 ms on an RTX PRO 6000, 28 ms on an M4 Max. 💡 Why it matters Jeff beats the much larger Jev on 5 of 6 classification benchmarks (79.1 vs. 83.0 overall) and hits 95.8% accuracy after a 30-minute fine-tune. No cloud GPUs, no closed-model outputs in training. ⚡ Our take If you're still calling GPT-4o for intent detection, you're burning money and latency. Jeff proves 0.8B parameters are enough — if you train them right.