Open-weight alignment and evaluation

Nous Research releases Arete-27B with open weights, training data and alignment evaluation

A Qwen3.8-27B supervised fine-tune exposes how post-training changes political response framing, but the reported improvements depend on the authors' own rubric and judge.

Open Model Weights published16 Nov 2026
Primary sourceNous Research — Arete-27B official technical report
Source published2026-11-06
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original atmospheric visualization of a 27-billion-parameter model represented by two diverging luminous neural pathways and a central evidence lens on a violet graphite background.
Original conceptual artwork about fine-tuning and evaluation methods; it does not represent a measured bias score or a political endorsement. AI-generated editorial illustration; not a documentary photograph.

Nous Research has released Arete-27B, an Apache-2.0-licensed supervised fine-tune of Alibaba's Qwen3.8-27B that studies how an additional training phase can alter responses to politically sensitive questions. The publisher presents the project as an effort to reinforce what it calls Western values in a Chinese-trained model. That framing is the developer's stated objective; it is not a neutral or independently verified definition of correctness. The scientific significance is the unusually inspectable training and evaluation process, including downloadable model weights, a training dataset and published benchmark materials. Primary source ↗

According to the model card, Nous trained the language-model weights on approximately 52.8 million tokens across two epochs and 1,202 optimizer steps, using text-only examples and keeping the vision encoder frozen. The project used an open teacher model and supervised fine-tuning rather than reinforcement learning or preference optimization. The resulting checkpoint remains a derivative of Qwen3.8-27B, not a new base architecture, and its full deployment, license and safety implications should be assessed at the actual published repository revision.

The central reported evaluation comprises 803 prompts spanning seven political and social categories. Nous says its internal rubric raised mean factual recall from 2.32 to 2.98 on a five-point scale, while the proportion of responses classified by its automated judge as adopting Chinese official framing fell from 22 percent to roughly 0.9 percent. The test construction, source selection, categories and judge criteria are part of the result, not an external standard of political neutrality. The benchmark has not been reproduced by Open Model Weights.

Nous also reports limited differences from its base checkpoint on selected general and agentic tasks. These developer measurements are useful leads for replication, not proof of unchanged capabilities in all workloads. Transparent access to weights, data and evaluation scripts lets other researchers probe the tradeoffs and challenge the assumptions behind the benchmark. Arete-27B illustrates the flexibility of open-weight post-training while underscoring that alignment is an empirical and often contested evaluation problem rather than a single universal score.

OMW REGISTRY WATCH

What this changes for the evidence layer

Existing curated Arete-27B model is Notable; do not raise it to Leading on Nous Research's own alignment benchmark. Separate a publisher-defined political-framing rubric from independently verified factual accuracy, checkpoint licensing and general capability.

PRIMARY SOURCE

Nous Research — Arete-27B official technical report

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗