Nous Research releases Arete-27B, a Western-values fine-tune of Qwen3.8-27B

Nous Research took Alibaba's Qwen3.8-27B model and fine-tuned it on about 52.8 million tokens, roughly 50,000 samples, using a 64k context window over two epochs and 1,202 training steps, while keeping the vision encoder frozen. The goal was to reduce propaganda-style answers and improve factual grounding without hurting general capability. On an 803-prompt evaluation set, the share of answers containing propaganda dropped from 22 percent to 1 percent, and the average fact score rose from 2.32 to 2.98 out of 5. The number of answers scoring zero fell sharply, from 359 to 216. General capability held steady: Hermes Bench came in at 65.60 percent, nearly identical to the base model's 65.77 percent, while Terminal-Bench-2.1 actually improved, moving from 67 percent to 70 percent. Nous released the model weights, the evaluation set, and the training data itself under the Apache 2.0 license, meaning anyone can inspect, reuse, or build on the work.

Why it matters

It is a concrete, measurable attempt to reduce bias in an open model without sacrificing performance, and a test case for how much fine-tuning alone can shift a base model's worldview.

Source: Nous Research