Not tested by us yet. Arete-27B ships as downloadable weights and no hosted provider serves it, so we could not call it through our harness. Everything below comes from the model card on Hugging Face and OpenRouter's public model list. We'll update this post when a hosted endpoint appears and we can run it against base Qwen3.8-27B on our own prompts.
Arete-27B is a 27B-parameter text model from Nous Research, a supervised fine-tune of Alibaba's Qwen3.8-27B under the Apache-2.0 license. Its goal is narrow: on politically sensitive topics, give factual answers instead of state-framed ones. Nous reports that the share of answers containing propaganda dropped from 22% for the base model to 0.9%, measured with an LLM judge on Nous's own 803-prompt set.
| Arete-27B | |
|---|---|
| Maker | Nous Research |
| Base model | Qwen3.8-27B (Alibaba) |
| Parameters | 27B, BF16 safetensors |
| License | Apache-2.0 |
| Modality | Text only (the vision encoder was frozen during training) |
| Training | Supervised fine-tuning only, 1,202 steps, about 52.8M tokens, 2 epochs |
| Training context | 64k tokens |
| Announced | 2026-11-06 |
| Hosted API | None found; no Inference Provider lists it |
What changed from the base model
Nous did not touch the architecture and did not use reinforcement learning or preference optimization. The recipe, per the card, is about 20,000 text-only prompts across seven value areas: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. Responses came from a teacher model and from the starting model.
The teacher is Inkling, which the authors call the best open model with a positive evaluation on these behaviors. They chose an open teacher to avoid terms-of-use problems with distillation. Inkling is the Thinking Machines model that also appears on OpenRouter as thinkingmachines/inkling, listed at $1.00 per million input tokens and $4.05 per million output tokens on the OpenRouter model list.
The card is frank about the limits: the authors say they addressed a root cause of CCP-framed answers but did not fix every viewpoint-related behavior.
The vendor-reported numbers
All figures below come from the model card. The propaganda evaluation uses 803 prompts, with Grok 4.7 as the judge and a 0 to 5 factual score.
| Model | Fact mean | Any propaganda |
|---|---|---|
| Inkling (teacher) | 3.96 | 0.3% |
| Arete-27B (thinking off) | 2.98 | 0.9% |
| Qwen3.8-27B (base, thinking off) | 2.32 | 22% |
On general ability, Nous reports Hermes Bench, an internal benchmark, at 65.77% for the base model and 65.60% for Arete, which the authors call noise. On Terminal-Bench-2.1 they report 67% for the base and 70% for Arete, run on their own hardware, and they say they do not claim that gain as a real improvement. We read it the same way: the fine-tune appears not to have damaged coding-agent ability, which is a smaller claim than "improved it."
Read the propaganda table with care:
- The judge is an LLM, and its rubric was refined on a 25-prompt pilot graded by hand.
- Arete's numbers were taken at temperature 0 while several comparison models ran at 0.6. At 0.6 the card gives Arete 2.88 and the base model 2.24, so the gap holds, but the settings are not identical.
- The biggest gain was in freedom of expression, and the benchmark was built by the same team that built the model.
Treat the 22% to 0.9% drop as a claim to reproduce, not a result.
Price and access on paper
You cannot rent Arete yet, so the nearest comparison is the base model. Qwen3.8-27B is on OpenRouter at $0.15 per million input tokens and $1.875 per million output tokens, with a 1,024,000-token context window, per the same model list. If Arete shows up on a router at a similar price, it would sit in the cheap tier. Today the only route is to download the weights and serve them yourself. The card lists 11 quantized derivatives and one fine-tune, and gives no hardware requirement beyond the 27B parameter count. For what self-hosting involves, our guide to open-source LLMs you can run on-prem walks through the options.
Who should care
- Teams serving users in politically sensitive domains (news, education, civic tools) who want a Qwen-class model with fewer state-framed answers, and who can afford to verify the claim on their own questions first.
- Researchers studying how post-training shifts a model's values. The card says this is an intended use, and the training dataset,
NousResearch/arete-data-pairs, is public. - Everyone else can skip it. If you care about coding or general chat, the evidence says Arete is about equal to the base model, not better, and the base is already hosted.
One caution on the premise. A model tuned toward one set of values is still a model with a viewpoint; the card itself describes the shift as toward Western values. Whether that is a feature depends on your users. For a quick refresher on how mixture-of-experts and dense models differ in serving cost, see what a mixture-of-experts model is.
What we'll test when we can
Once a hosted endpoint exists, we'll run Arete-27B against base Qwen3.8-27B on 10 to 15 politically sensitive questions plus a coding set, judged with swapped answer order, and report cost and latency alongside.