What is Arete-27B?
Arete-27B is a fine-tune of Alibaba's Qwen3.8-27B, released by Nous Research on 6 November 2026. Nous says it post-trained the model "to reinforce western values" after finding refusals, deflections and answers framed in the Chinese Communist Party's terms on politically sensitive topics in the base model. It targets seven areas: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy.
The method is supervised fine-tuning only, with no reinforcement learning and no architecture change: about 52.8 million tokens over two epochs, with answers taken from Thinking Machines' open Inkling model as the teacher. Nous published the 803-prompt evaluation it built, Propaganda_Eval, and the training pairs alongside the weights. The vision encoder was frozen, so image input is Qwen3.8's, untouched.
The weights are under Apache 2.0, like the base model. Nous offers no hosted API; community quantisations, the most popular a GGUF set from bartowski, appeared within days.
Is Arete-27B worth it?
Arete's interest lies in what it proves more than in what it adds. On Nous's own test of 803 prompts drawn from English-language statutes, news and research, the share of answers that adopt the Chinese state's account fell from 22% to 0.9%, and the average fact score rose from 2.32 to 2.98 out of 5. The largest change was on freedom of expression, where the state-framing rate fell from 60% to 1%. Nous's conclusion is that this kind of bias is applied late and thinly, so it can be shifted with a small amount of supervision.
The numbers need care. Nous wrote the benchmark, chose the facts, and used Grok 4.7 as the judge, tuned on a 25-prompt pilot it graded by hand. Arete and the base model ran at temperature 0 while the comparison models ran at 0.6; at 0.6 Arete scores 2.88. The Decoder notes that Inkling, the teacher, was itself trained partly on output from a Chinese model, and at least one critic on X argues that Nous has swapped one set of values for another, a point the model card does not dispute: it frames the work as giving users a choice of viewpoint. The method works in any direction.
General ability looks unchanged. On Nous's internal Hermes Bench it scores 65.60% against 65.77% for the base, which Nous calls noise, and it reports 70% against 67% on Terminal-Bench 2.1 while declining credit for the gain. There is no independent leaderboard score.
Our take: for a team that already likes Qwen3.8-27B but needs it to answer questions on China, surveillance or democracy without the official line, Arete is a drop-in swap with the same size, licence and tooling. For anyone else it is the same model, and the published eval is the more lasting contribution.
Where it earns its place
A newsroom data team at a European publisher
Run a local assistant that summarises reporting on China, Hong Kong and Taiwan without echoing official language
Swap Arete into an existing Qwen3.8-27B setup; Nous's eval shows the state-framing rate on freedom of expression falling from 60% to 1%. Still check facts: the average fact score is 2.98 out of 5.
An AI safety researcher studying how post-training shifts values
Measure how much a narrow fine-tune moves a model's political framing
Nous published the 803-prompt Propaganda_Eval with judge prompt and scripts, plus the training pairs. Rerun it with another judge, or at matched temperatures, to test the 22% to 0.9% claim.
A procurement lead at a US public agency wary of Chinese-trained models
Deploy a capable open model on internal hardware
Arete keeps Qwen3.8-27B's weights layout and Apache 2.0 licence, so the BF16 file of about 55 GB runs where the base does. The base model's origin is unchanged, which some policies will still count.
Before you commit
- Setup
- Download the 18-file BF16 checkpoint from Hugging Face and serve it as you would Qwen3.8-27B; community GGUF and MLX builds run in llama.cpp and on Apple silicon.
- Integrations
- Uses Qwen3.8-27B's chat template, so vLLM, SGLang, llama.cpp and Transformers load it unchanged. Nous evaluated with thinking off; a thinking-on score of 3.07 is from two of five runs.
- Data and privacy
- Self-hosted, so prompts stay on your machine. Apache 2.0 allows commercial use and modification; the evaluation set is published under MIT.
- Cost at scale
- Weights are free. About 27.4 billion parameters, roughly 55 GB in BF16, so full precision needs an 80 GB GPU; community 4-bit GGUF and NVFP4 builds cut that to a single consumer card's range.
- Limits
- Fine-tuned at 64K context, against the base's 262K. Nous does not claim to have removed every effect of the base model's viewpoint, and the multi-token-prediction head was disabled during training.
What’s good
- State-framed answers fall from 22% to 0.9% on Nous's eval
- No measured capability loss on Hermes Bench
- Evaluation set and training pairs published
- Apache 2.0, same size and tooling as Qwen3.8-27B
What’s not
- Nous built the benchmark, picked the judge and ran every test
- Different decoding settings for Arete than for the comparison models
- Fact score of 2.98 out of 5 still trails its teacher's 3.96
- No hosted API and no independent leaderboard score
Best for
Teams running Qwen3.8-27B locally who need it to answer politically sensitive questions without Chinese state framing.
Arete-27B alternatives
Arete-27B: common questions
- Is Arete-27B worth it?
- Arete-27B shows that a Chinese model's political framing comes off with one modest fine-tune, but it is a Qwen3.8-27B with a different worldview, not a better Qwen. We rate it 3.7 out of 5.
- How much does Arete-27B cost?
- Free Apache 2.0 weights. No first-party API; you pay for your own hardware.
- What are the downsides of Arete-27B?
- Nous built the benchmark, picked the judge and ran every test; different decoding settings for Arete than for the comparison models; fact score of 2.98 out of 5 still trails its teacher's 3.96; no hosted API and no independent leaderboard score.
- Who is Arete-27B best for?
- Teams running Qwen3.8-27B locally who need it to answer politically sensitive questions without Chinese state framing.
- How much does Arete-27B cost for a team or in production?
- Weights are free. About 27.4 billion parameters, roughly 55 GB in BF16, so full precision needs an 80 GB GPU; community 4-bit GGUF and NVFP4 builds cut that to a single consumer card's range.
- Is Arete-27B safe for company data?
- Self-hosted, so prompts stay on your machine. Apache 2.0 allows commercial use and modification; the evaluation set is published under MIT.
- What are the best Arete-27B alternatives?
- The closest alternatives we have reviewed are Qwen3.8-27B, Inkling, Gemma 4 and Mistral Large 4.
Answers checked Nov 11, 2026.