

Jev-Style 2B v4
#16 in Small & Edge ModelsChaoliang Yan · v4 · 2b · since 2026-10-27 · 2× · last seen Nov 14, 2026
32
Momentum
Jev-Style 2B v4 is an open-weight 2B decision model (base: Qwen3.5-2B with vision tower) by Chaoliang Yan that returns a calibrated probability for every option in one pass. It supports text and images (PyTorch build; MLX and GGUF builds are text-only), has a 32,768-token input window, and is licensed under Apache-2.0. Sources stress independence from TypeSafe; the model was trained in part on soft labels distilled from TypeSafe Jev 1.14 via its public API.
Momentum trend
17.08.15.11.
Features
| Key Benchmark (%) | JevBench v1.4.1: 74.9% (public items); Decision Index 0.3 Public: 44.37 |
| Context Window (Tokens) | 32,768 tokens (total input) |
| License | Apache-2.0 |
| Multimodality | Text and images (PyTorch build only); base Qwen3.5-2B with vision tower |
| Platform | Local (PyTorch: text+image; MLX and GGUF: text-only); hosted API: text-only, 8,192-token context |
| Release Date | October 27, 2026 |