GenBase Rhythm open-sources the decision model NeoHorse-Jev; the official summary describes it as ranking first in the average score of the open-source Jev comprehensive evaluation

Article

GenBase Rhythm open-sources the decision model NeoHorse-Jev, ranking first in the open-source Jev comprehensive evaluation

Beating AI news: GenBase Rhythm, founded by Wang Yunhe, former head of Huawei’s Pangu large model, has open-sourced NeoHorse-Jev-, turning its own NeoHorse-1 into a Jev-type decision model. It does not need to generate answers character by character; instead, it directly selects from several candidates, makes binary judgments, or assigns scores. It is mainly used by Agents to select tools, perform routing, and decide on the next action.

The official summary covers 6 text evaluations, including JevBench, Kev, and OpenJev. NeoHorse-Jev-’s equally weighted average score was 77.70, higher than the 75.67 of the -parameter Open-Jev and also above the 74.25 of Kev-. However, this composite score was compiled independently by GenBase Rhythm. At present, only 4 open-source models have complete results across all six evaluations, and NeoHorse-Jev achieved the highest score in only two of them.

It is based on NeoHorse-1 and directly calculates the probabilities of candidate answers through prefill-only inference, eliminating the process of generating text character by character used by ordinary large models. In a test involving 1158 text requests, the server’s average decision latency per request was 32 under concurrent load, and throughput reached 74.3ms；256 requests per second under concurrent load.

The model supports text and single-image inputs. Its weights, inference code, and deployment examples have all been made available. It can run through vLLM, SGLang, or a local Runtime, under the Apache 347.05 License.

Original link 2.0 https://m.theblockbeats.info/flash/368930

Key points

The report describes NeoHorse-Jev- as a Jev-type decision model built on NeoHorse-1.
The model does not generate answers character by character; instead, it selects, judges, or scores candidates. It can be used by Agents to select tools, perform routing, and decide on the next action.
The official summary 6 text evaluations and states that NeoHorse-Jev-’s equally weighted average score was 77.70, while Open-Jev scored 75.67, and Kev- scored 74.25。
At present, only 4 open-source models have complete results across all six evaluations, and NeoHorse-Jev achieved the highest score in only two of them.
In a test involving 1158 text requests,32 under concurrent load, the server’s average decision latency per request was 74.3ms；256 under concurrent load, throughput was 347.05 requests per second.
The model supports text and single-image inputs; its weights, inference code, and deployment examples have all been made available.
The model can run through vLLM, SGLang, or a local Runtime, under the Apache 2.0 License.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a model specifically designed to help Agents “make decisions.” Ordinary large models often generate answers character by character, whereas this model directly calculates the probabilities of several candidate answers and then selects, judges, or scores them, making it suitable for tool selection and task routing.

Why it matters to readers

The release of the open weights, inference code, and deployment examples enables developers to try and deploy this type of decision model themselves. The test data disclosed in the report also shows its latency and throughput performance in concurrent decision-making scenarios.

Beginners can think of it as an Agent’s “selector”: when a system has multiple tools or possible next actions, the model evaluates the candidates rather than composing an entire answer.

Risks and unknowns

The composite score was compiled independently by GenBase Rhythm, and the original text does not establish whether it underwent independent verification.
There are currently only 4 open-source models with complete results across all six evaluations, limiting the scope of the comparison.
The original text does not disclose the test environment, hardware configuration, or specific measurement methodology.
The original text does not provide accuracy, stability, or image-input performance data from actual community deployments.

GenBase Rhythm open-sources the decision model NeoHorse-Jev; the official summary describes it as ranking first in the average score of the open-source Jev comprehensive evaluation

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center