Nous Research's Arete-27B Cuts Propaganda Answers From 22% to 0.9% on Its Own Test

Nous Research published Arete-27B, a fine-tune of Alibaba's Qwen3.8-27B that its model card says was post-trained "to reinforce western values on a Chinese-trained model", and reports that answers containing propaganda fell from 22% to 0.9% on its own evaluation. The weights are on Hugging Face under Apache-2.0, with 5,545 downloads and 512 likes in the last month as of November 13.

The model card credits Nous Research as a team and names no individual authors. It lists the base as Qwen/Qwen3.8-27B, a 27-billion-parameter model, and says the training method was "supervised fine-tuning only", with no reinforcement learning, preference optimization or architectural changes.

How it was trained

Nous used roughly 20,000 generated prompts across seven value categories, with teacher responses from a model the card calls Inkling. Training ran 52.8 million tokens over 2 epochs and 1,202 steps at batch size 48, using AdamW, with a 64k training context. The data is published as NousResearch/arete-data-pairs.

The card says the Propaganda_Eval dataset it links is a held-out evaluation set and not training data. No inference provider hosts the model yet.

What the card reports

The headline measurement is a propaganda evaluation Nous built, with a mean factuality score on a 0 to 5 scale. The card also reports an internal Hermes Bench and Terminal-Bench 2.1. All figures are vendor-supplied, and the card says decoding differs across rows: the propaganda comparison used temperature 0 with thinking off, while other models in its table used temperature 0.6.

Benchmark	Qwen3.8-27B (Nous table)	Arete-27B
Propaganda eval, fact mean (0 to 5)	2.32	2.98
Answers with propaganda	22%	0.9%
Hermes Bench (internal)	65.77%	65.60%
Terminal-Bench 2.1	67%	70%

1. Qwen3.8-27B22 %
2. Arete-27B0.9 %

Source: Arete-27B model card, Hugging Face, accessed 2026-11-13

Where the numbers disagree

Qwen's own Qwen3.8-27B card reports 73.0 on Terminal Bench 2.1 for the base model. Nous's table gives the base 67% and Arete 70%, so Arete sits below the figure Qwen publishes for its own checkpoint. Qwen's card also lists 61.7 on SWE-bench Pro and 84.3 on OSWorld-Verified, neither of which Nous reports.

The card itself says the authors "do not take credit" for the Terminal-Bench gain, and reports "no meaningful capability regression" on the benchmarks it ran. It also says the work does not "remedy all behaviors" tied to the Chinese state's viewpoint.

The evaluation that carries the headline number was built and scored by the same group that trained the model, which is the main reason to wait for an outside run. Both cards list Apache-2.0, and the Qwen card lists a 262,144-token native context against Arete's 64k training context.

Context on Chinese base models

The Terminal reported that six Chinese models read as Blackwell-trained in a seed fingerprint report, and covered Runway's Apache 2.0 Praxis-1 release where a third-party leaderboard supplied the ranking.

The card states no release date for Arete-27B.