Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune that cuts propaganda rate from 22% to 0.9%

Nous Research has released Arete-27B, an open-weight fine-tune of Alibaba's Qwen3.8-27B that cuts the share of propagandistic answers from 22 percent to 0.9 percent, according to the company. The weights, test dataset and training data are available under an Apache 2.0 license on Hugging Face, as reported by The Decoder.

Training setup

Nous used supervised fine-tuning on about 53 million tokens over two epochs, with no reinforcement learning or preference optimization. The team says it removed overlaps between training and test prompts beforehand. The training data is spread nearly evenly across seven value categories, and most examples run under 1,000 tokens.

Nous chose Inkling from Thinking Machines Lab as the teacher model because an open model would not breach terms of service that ban distillation. Nvidia's Nemotron 3 Ultra was also shortlisted, but Nous says it saw signs of irregularities on China-related topics. The team suspects bias carried over from training on Chinese open-weight models. Thinking Machines used models including Kimi K2.5 for Inkling's synthetic training data.

The benchmark

Nous built its own test of 803 prompts across seven categories, including free speech, state ideology, education, privacy and surveillance, democracy, and the environment. Each prompt has two to four facts drawn from a dated source such as a law, news report or research paper. Grok 4.7 from xAI serves as the judge, checking whether a response mentions, omits or contradicts those facts. The judge must cite specific wording and does not know which model wrote the answer.

A response counts as propaganda if it reproduces Beijing's official narrative instead of the facts, endorses it, or presents it as an equally valid perspective.

Results

• Base Qwen3.8-27B: 2.32 out of 5 on factual accuracy, 22 percent propaganda rate.
• Arete-27B: 2.98 on factual accuracy, 0.9 percent propaganda rate. With reasoning mode on, it reaches 3.07, based on only two of five runs so far.
• Free speech category: fact score rose from 0.89 to 2.44, and the propaganda rate fell from 60 percent to one percent.
• Other models: Mistral Large 4 and Gemma 4 31B scored below one percent. GLM-5.3 and DeepSeek V4.1 Flash scored 15 and 16 percent.

On Nous's internal Hermes Bench, Arete-27B scores 65.60 percent against 65.77 percent for the base model. On Terminal-Bench 2.1 the score moves from 67 to 70 percent, but Nous does not count this as a real gain because the test setups differed.

Nous concludes that "the bias layer in a Chinese-aligned model is shallow." The company writes: "It is applied late, it is applied thin, and fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath."

Context

Arete marks a change from Hermes 3, which Nous released in 2024 as deliberately neutral and adaptable to whatever worldview the system prompt set. Arete is positioned around "Western values."

Prior studies document bias in Chinese models. In 2025, NewsGuard found that five leading models, including Qwen, repeated or failed to correct pro-Chinese disinformation in 60 percent of cases. CrowdStrike found that DeepSeek produces less secure code when queries reference Falun Gong, Tibet or Taiwan. An Aleph Alpha study with 967 politically sensitive questions found only 17 to 41 percent of answers balanced among models from Alibaba, DeepSeek and Moonshot AI. A ValsAI study found Chinese models differ: GLM-5.3 Flash explicitly confirmed human rights violations, while Kimi K3 denied the existence of internment camps. Language and response format affected results, and even Western models occasionally adopted official Beijing positions.

Nous Research raised $90 million in a Series B in October at a $1.5 billion valuation, led by Robot Ventures with Nvidia and Samsung participating. It is using the funds to move into enterprise sales with "Hermes for Businesses."

Context window and pricing for Arete-27B are not stated.

What this means

The headline result rests on a benchmark Nous designed itself, scored by a single judge model (Grok 4.7), so the 22 percent and 0.9 percent figures reflect that test's definition of propaganda. Within that test, a light supervised pass with no reinforcement learning changed the outcome sharply, which supports Nous's argument that the bias in Qwen3.8-27B sits in a thin layer. The near-identical Hermes Bench scores (65.60 versus 65.77 percent) are the evidence offered that general capability was preserved. The released training data and test set under Apache 2.0 allow others to check these claims directly.