Nous Research has introduced Arete-27B, a version modified through fine-tuning on the base model Qwen3.8-27B. The development was released on November 8, 2026, under the terms of the Apache 2.0 open-source license, allowing for inspection, use, and integration within research and technical production environments.
The primary objective of this iteration is to limit the generation of ideological or biased content in automated responses. In the logs shared by the Nous Research team, the rate of responses identified as propaganda experienced a substantial decrease compared to the records of the base model.
Results in the 803-instruction evaluation
The behavior of Arete-27B was measured through a specific evaluation framework designed by Nous Research consisting of 803 test instructions. In this examination, the proportion of responses classified as propaganda stood at 0.9%, compared to the 22% originally recorded by Qwen3.8-27B.
The reduction represents a change of more than 21 percentage points in the system's propensity to issue doctrinaire statements in response to complex requests.
Performance stability on Hermes Bench
One of the common issues in alignment and bias mitigation processes is the degradation of model capabilities. On this occasion, the evaluation analyzed the impact of the filter using the Hermes Bench test.
According to data reported by Nous Research, the score obtained by Arete-27B on Hermes Bench remained exactly flat at 65.60%. The fine-tuning succeeded in suppressing propaganda outputs without inducing a loss of performance in standardized evaluation tasks.
Availability under Apache 2.0 license
The adoption of the permissive Apache 2.0 license makes Arete-27B available to the community for use without software royalty costs.
