Arete-27B

Features

• Fewer state-framed answers. On Nous's own 803-prompt evaluation, answers that adopt the Chinese state's account fell from 22% for Qwen3.8-27B to 0.9%, and the fact score rose from 2.32 to 2.98 out of 5. Official source 1 ↗ · Official source 2 ↗
• Seven value categories. Training targeted freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy and environmental policy. The biggest change was in freedom of expression, where propaganda answers fell from 60% to 1%. Official source 1 ↗
• Capabilities kept. Nous reports 65.60% on its internal Hermes Bench against 65.77% for the base model, which it calls noise, and 70% against 67% on Terminal-Bench-2.1. Official source 1 ↗
• Open training data and eval. The training pairs (NousResearch/arete-data-pairs) and the evaluation, with every prompt, its source article, the judge prompt and grading scripts, are published so you can test this or any other model. Official source 1 ↗ · Official source 2 ↗
• Small, documented fine-tune. Supervised fine-tuning only, on about 52.8 million tokens at 64k context over two epochs, with the full optimizer settings in the model card. No RL, no preference optimization and no architecture change. Official source 1 ↗
• Ranked against other open models. In the same eval, the teacher Inkling scored 3.96 with 0.3% propaganda; GLM 5.3 scored 3.10 but with 15% of answers adopting the state line. Official source 1 ↗

Pricing & free limits

Billing details

Billing details were not established by the checked official sources.

Free access & limits

Free access and its limits were not established by the checked official sources.

Best for & limitations

Running a Qwen-class local assistant or agent that answers questions on Chinese politics, history and governance without adopting the state's framing. Official source 1 ↗ · Official source 2 ↗

Research into how post-training changes a model's values, using the published training pairs and evaluation. Official source 1 ↗ · Official source 2 ↗

Benchmarking other open models for state-aligned framing with the released 803-prompt eval and grading scripts. Official source 1 ↗ · Official source 2 ↗

• A deliberate viewpoint change. This is a value-alignment fine-tune: Nous describes it as post-training "to reinforce western values" and chose a teacher whose answers matched its own targets. Whether that suits you depends on the viewpoint you want. Official source 1 ↗
• Evaluation is Nous's own. The headline numbers come from Nous's own evaluation graded by an LLM judge (Grok 4.7), with different decoding settings across models; the thinking-on score is provisional from 2 of 5 runs. Official source 1 ↗ · Official source 2 ↗
• Not all behavior addressed. Nous says it has not remedied every behavior tied to the base model's viewpoint, only a potential root cause. Official source 1 ↗
• Capability claims are vendor-run. Hermes Bench is an internal benchmark, and Nous runs both models on its own hardware; it does not take credit for the Terminal-Bench gain because configurations differ from official results. Official source 1 ↗ · Official source 2 ↗
• Text-only fine-tune. Training data was text-only and the vision encoder was frozen; the multi-token-prediction head was disabled during training. Official source 1 ↗
• Large for a laptop. At 27B parameters it needs a large GPU or heavy quantization to run locally, and Nous ships the weights without its own hosted API. Official source 1 ↗