News · 2026-11-07

Nous releases Arete-27B to change responses to political prompts

Nous Research has released Arete-27B, a downloadable fine-tune of Qwen3.8-27B intended to change how the model answers politically sensitive questions. Nous reports improvements on its own 803-prompt evaluation, while independent validation of the framing and results remains open.

Key facts

• The official model card identifies the Qwen3.8-27B base and Apache-2.0 license.
• The original repository contains about 54.7 GB of safetensors weights.
• Nous evaluated the model on 803 prompts in its own test.
• The reviewed sources do not state a minimum or recommended video-memory requirement.

Nous calls its desired alternative “Western values,” a political framing that should be attributed to the lab rather than treated as a neutral scientific result. Its announcement describes reducing refusals and particular state-aligned framing on sensitive subjects. The model is a fine-tune: it starts with an existing model and trains further on selected examples, like revising a reference book’s editorial habits without rebuilding its printing press.

The lab says it used supervised training, not reinforcement learning, and did not change the underlying architecture. It published a paired dataset and evaluation materials, making closer inspection possible. The currently visible dataset is larger than one prompt-generation figure in the announcement; the reviewed materials do not establish that every released row trained the published checkpoint.

The original model files occupy about 54.7 GB on disk. A community conversion offers a roughly 17.2 GB Q4_K_M file. Disk space is not the same as graphics memory: the reviewed primary sources do not specify a minimum or recommended GPU memory requirement for the checkpoint, so no VRAM figure can responsibly be inferred from the download size. Quantization and offloading can change runtime needs and behavior.

Nous designed the evaluation and its criteria, including which facts should be present and what constitutes adoption of a particular framing. A model can score better by that definition without becoming universally unbiased. Its own general-capability tests provide limited checks, not a guarantee that other capabilities and languages are unaffected.

Early model discussions include objections to the values framing and requests for deployment features. Those comments are not a representative survey. The strongest next evidence would be independent reruns, examination of factual mistakes, and comparisons across viewpoints and real tasks.

The training method matters because it narrows what the release demonstrates. Nous says it continued supervised training on selected examples while leaving the architecture unchanged. The lab also says the vision encoder was frozen and that its training examples were text-only. A change in political answers therefore does not establish a new pretrained family or a general improvement to every task supported by the base model.

The lab's evaluation asks whether answers contain selected facts and whether they adopt a particular official framing. Nous says it uses a model judge and applies a penalty when a required fact is contradicted. A judge can help evaluate hundreds of responses consistently, but it introduces another model's judgments into the result. Human review, alternative judges and transparent examples would help show which changes reflect factual improvement and which reflect the chosen framing. The site's lesson on model judges covers this dependency.

The published training and evaluation datasets make such tests possible, but publication is not the same as a fully reproducible training run. The dataset card describes accepted and rejected examples, and the visible dataset contains more material than one earlier prompt count in the announcement. The reviewed documents do not pin every released row to the checkpoint. Readers should not assume that every public example was used in its training or that no errors remain in the accepted subset.

The official weights and the community quantization are also separate artifacts. A smaller file can make local experimentation easier, but the conversion may affect output quality and may use different runtime formats. The community card provides run instructions; the reviewed materials do not give a verified minimum graphics-memory figure for either setup. Anyone planning hardware should obtain a stated or measured configuration for their intended runtime instead of interpreting disk size as memory use.

Primary source, verified: read the paper →

Ground Truth. (2026, November 7). Nous releases Arete-27B to change responses to political prompts. Ground Truth. https://groundtruth.day/news/nous-arete-27b-political-finetune.html

@misc{groundtruth:nous-arete-27b-political-finetune,
  title  = {Nous releases Arete-27B to change responses to political prompts},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/nous-arete-27b-political-finetune.html}
}

Topics: open-weights · models · evaluation