Jeff v1.1 expands local classification to 254 choices

Jeff is an independent collection of small decision models using Jev’s request format. Give it a situation and named options, and it returns probabilities rather than generated prose. Version 1.1 expands its Qwen-based models from 26 to 254 choices. The project reports better long-list results, while also disclosing a small decline on its larger model’s aggregate benchmark.

Jeff project · Sep 29, 2026

Why it matters

Local routing and classification can be evaluated without a general text-generation workflow.

More context

Try the serving-only Python installation and a published checkpoint on CPU, CUDA or supported Apple Silicon configurations. Measure accuracy on your own labels. Jeff is not affiliated with TypeSafe, and the expanded option limit does not apply to its Gemma variant.

Keep in mind

This is an early independent project. Benchmark and calibration claims are maintainer-reported, not reproduced by 1min; request compatibility does not establish equivalent reasoning quality to Jev.

Original read