Primitive Rhythm Open-Sources 4B Decision Model NeoHorse-Jev, Jev Ranks First in Comprehensive Evaluation

Beating AI News Flash: NeoHorse-Jev-4B, open-sourced by YuanLvDong, founded by former Huawei Pangu Large Model head Wang Yunhe, turns its own NeoHorse-1-4B into a Jev-style decision model. It does not generate answers word by word, but directly selects among several candidates, judges true or false, or assigns scores, mainly used for Agents to choose tools, perform routing, and decide the next action.

Officials summarized 6 text evaluation sets including JevBench, Kev, and OpenJev. NeoHorse-Jev-4B's equal-weight average score is 77.70, higher than 9B-parameter Open-Jev's 75.67, and also exceeds Kev-4B's 74.25. However, this composite score was summarized by YuanLvDong itself, and currently only 4 open-source models have complete results across all six items, while NeoHorse-Jev also only achieved the highest score in two of those groups.

It is based on NeoHorse-1-4B and directly calculates the probability of candidate answers through prefill-only inference, eliminating the process of ordinary large models generating text word by word. In a test of 1,158 text requests, at 32 concurrency the server averaged 74.3ms per decision; at 256 concurrency, throughput reached 347.05 requests per second.

The model supports text and single-image input. Weights, inference code, and deployment examples are all open, and it can run through vLLM, SGLang, or a local Runtime, under the Apache 2.0 license.