Arete-27B Recasts Qwen3.8’s Political Defaults

Arete-27B is less a capability breakthrough than a clear demonstration that open weights let labs retune a model’s political defaults. Its bigger contribution may be making value alignment more explicit—and forcing developers to ask whose rubric defines “neutral” behavior.

• –Uses supervised fine-tuning only, with Inkling as the teacher and no reinforcement learning or architectural changes.
• –Targets freedom of expression, state ideology, markets, education, privacy, democracy, and environmental policy.
• –Nous reports no meaningful regression on general and agentic benchmarks, but independent replication is still needed.
• –The release highlights open-weight models as adaptable foundations rather than fixed ideological products.
• –Developers should evaluate Arete against their own safety, factuality, and political-bias requirements before deployment.