Jeff: local zero-shot classification in 22 ms with a 0.8B model

The independent GitHub project Jeff performs zero-shot classification with a Qwen3.5-0.8B model in 22 ms on an RTX PRO 6000 and 28 ms on an Apple M4 Max. After fine-tuning, its average score rose from 45.3 to 79.1, and it hit 96.4 on Financial PhraseBank versus 77.0 for the larger Jev.

• Jeff-Qwen3.5-0.8B responds in 22 ms on RTX PRO 6000 and 28 ms on M4 Max
• Fine-tuning lifted the average score from 45.3 to 79.1 across five benchmarks
• On Financial PhraseBank Jeff scored 96.4 versus 77.0 for Jev
• Training runs locally: 2 hours for 0.8B and 3.5 hours for 2B on one GPU