Jiyuan Rhythm has open-sourced the 4B parameter decision model NeoHorse-Jev, which ranks first in comprehensive evaluations of open-source Jev models.

Beating AI Express News: NeoHorse-Jev-4B, an open-source decision-making model from Jiyuan Ludong (founded by Wang Yunhe, former head of Huawei’s Pangu large model), is built on the company’s NeoHorse-1-4B. Unlike standard large models that generate answers token-by-token, this model selects from multiple candidates, judges correctness, or assigns scores, primarily for agent tool selection, routing, and next-step action decisions. The developer aggregated results from six text benchmarks including JevBench, Kev, and OpenJev. NeoHorse-Jev-4B scored an equal-weight average of 77.70, outperforming the 9B-parameter Open-Jev (75.67) and Kev-4B (74.25). Note that the composite score was compiled by Jiyuan Ludong itself; currently only four open-source models have complete results across all six benchmarks, and NeoHorse-Jev only took top scores in two of them. Built on NeoHorse-1-4B, it uses prefill-only inference to directly calculate candidate answer probabilities, eliminating the token-by-token generation step common in standard LLMs. In tests with 1,158 text requests, the server recorded an average decision latency of 74.3ms at 32 concurrent connections, and a throughput of 347.05 requests per second at 256 concurrent connections. The model supports text and single-image inputs. Its weights, inference code, and deployment examples are all open-source, runnable via vLLM, SGLang, or local Runtime, and licensed under Apache 2.0.