An independent developer releases Jev-Style 2B Decision v3, a 2B model with a free hosted API

A Hugging Face user, chaoliang yan, published Jev-Style-2B-Decision-v3 on 27 September 2026, by repository creation date. The card calls it an independent project, not affiliated with TypeSafe AI, Jev, Laya or the Qwen team. It is a 2B decision model: give it a state and typed questions (choice, yes/no or score) and it returns a calibrated probability for each option in one pass. It does not generate text or take actions.

It has 1.88B parameters, is fully fine-tuned from Qwen/Qwen3.5-2B on a reduced 60M-token pool, and is licensed Apache-2.0. GGUF and MLX builds exist. Stock llama.cpp, Ollama and LM Studio cannot produce the decision scores; the shipped runtimes are needed. The input budget is 25,600 tokens, and the free hosted API caps input at 8,000.

The card reports 73.6% (170 of 231) on public JevBench v1.4.1 items, with a 95% interval of 67.6 to 78.9, and says this is not an official leaderboard entry. Jev 1.13 scores 86.6% on the same items. Fourteen benchmarks were in the training pool, so results on them are not zero-shot.

The card suggests sending low-confidence answers to a person or a larger model.

A modest model with a candid card. It is a tool for tinkering.

Sources

• https://huggingface.co/chaoliang yan/Jev-Style-2B-Decision-v3
• https://github.com/lawrence3699/jev-style

Editor's first pass (AI-assisted)

• Real 3.5: Weights, GGUF and MLX builds and a free hosted API with an 8,000-token cap.
• Useful 2.5: A small model that trails the one it imitates by 13 points on the same items.
• Notable 2: A hobby-scale project.
• Novel 2.5: Imitates an existing API.