The boundary between accepting a model's values and building one from scratch is more fluid than the debate around Large Language Models often suggests. Arete-27B, published by Nous Research, starts from Alibaba's Qwen3.8-27B and uses supervised fine-tuning alone to shift its orientation across seven politically sensitive areas: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. There are no architectural changes, no reinforcement learning, and no preference optimization. The work relies on text-only data and a 64k token training context.

The team separated two levels that are often conflated: the ability to cite verifiable sources and the framing through which an answer adopts the Chinese state's perspective. On that distinction, Arete-27B raised its fact score from 2.32 to 2.98 out of 5 and cut answers containing propaganda from 22% to 0.9 percent. The improvement held across all seven categories, with no meaningful regression on general-capability and agentic benchmarks. In other words, the model's competence was not sacrificed to correct its political tilt.

Choosing SFT alone has a precise technical value. Compared to reinforcement learning or preference optimization pipelines, supervised fine-tuning requires fewer steps and can be replicated by teams with limited resources, including in a self-hosted environment. For those running open source models inside their own systems, this reduces dependence on complex infrastructure and allows behavior changes without revisiting the architecture. Arete-27B does not claim to have fixed all behaviors inherited from the Chinese model, but it shows that a targeted intervention on training data can measurably change how answers are framed.

There is also a structural signal. Open source models become raw materials that labs can shape toward different preferences: the values a model ships with do not have to be the only ones available to its users. This shifts some power from the vendor to those with curated datasets and post-training skills. Companies evaluating on-premise deployment no longer have to choose only between accepting a model aligned elsewhere or undertaking full training: they can operate on an intermediate layer, with lower costs and faster timelines. For those evaluating on-premise deployment, there are trade-offs between fine-tuning costs and behavior control; AI-RADAR provides analytical frameworks at /llm-onpremise to assess them.

For organizations subject to sovereignty requirements or local regulations, on-premise fine-tuning intersects with data control: datasets used for realignment do not have to leave corporate boundaries. Arete-27B offers a concrete example of how an existing model can be reoriented without relying on external services. The fact remains that the fact score, while improved, stays below 3 points out of 5: propaganda collapsed, but answer reliability is not yet a strength. The demonstration is not that the values problem is solved, but that post-training can work as a surgical lever, at least to shift political framing without starting from scratch.