Arete-27B
Nous Research's retrained Qwen3.8-27B that changes how it answers politically sensitive questions, Apache 2.0, with open data and eval.
In short. Arete-27B is Alibaba's Qwen3.8-27B retrained by Nous Research to give fewer refusals and fewer answers framed in the Chinese government's terms on politically sensitive topics. Nous describes this as reinforcing Western values. The weights are free under Apache 2.0, and the test results come from Nous itself.
Summary
Arete-27B is Nous Research's post-training of Alibaba's Qwen3.8-27B. Nous says it saw refusals, deflections and answers framed in the Chinese state's terms on politically sensitive topics in the base model and trained to change that, which its blog post describes as reinforcing Western values on a Chinese-trained model. The method is plain supervised fine-tuning, with no reinforcement learning, preference optimization or architecture change: about 52.8 million tokens of text at 64K context over two epochs, using answers from a teacher model, the open Inkling, across seven topic areas such as freedom of expression, state ideology, democracy and data privacy. Nous says it picked an open teacher to avoid the terms-of-use limits on distilling from closed models.
On Nous's own evaluation of 803 prompts, which scores factual recall from 0 to 5 and whether an answer adopts the Chinese state's account as its own, with Grok 4.7 as the judge, Arete-27B with thinking off raised its fact score from 2.32 to 2.98 and cut answers containing propaganda from 22% to 0.9%. Nous says the gain held in all seven categories, and that Hermes Bench, its internal general-capability test, moved from 65.77% to 65.60%, which it reads as noise. On Terminal-Bench 2.1 it measured 70% against 67% for the base model, but says it takes no credit for that difference. The comparison table also lists other models, for example Mistral Large 4 at 3.65 and Nemotron 3 Ultra at 3.22 on the fact score, and Nous notes that the decoding settings are not identical across all of them. These are the developer's own results, built around its own goal, and we have not reproduced them.
The weights use Apache 2.0, the same license as the base model, and Nous published the evaluation prompts and the training data pairs as Hugging Face datasets. The checkpoint keeps Qwen's vision encoder frozen, but Nous trained and evaluated on text only. The repository first appeared on Hugging Face on 6 November 2026 and, in the 30 days to 11 November 2026, Hugging Face counted about 4,000 downloads.
Good for
- Researchers studying how a small amount of supervised fine-tuning changes a model's values and behavior, with open weights, data and evaluation
- General assistant and agent work where you want the framing on politically sensitive topics that Nous trained for
- Side-by-side tests against Qwen3.8-27B, which has the same size and license and differs only by this fine-tune
Watch out for
- The model is tuned toward a viewpoint its developer chose, which Nous itself calls Western values, so test it on the topics that matter to you before relying on it
- The evaluation, its seven categories and its LLM judge are all Nous's own, and the comparison table mixes decoding settings
- Nous trained on text only with the vision encoder frozen and at 64K context, so treat image, video and very long inputs as untested even though the config allows 262,144 tokens
- Nous itself says it did not remedy every behavior associated with the base model's viewpoint
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the 27B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| 27B BF16 | 27.4B | 63 GB | 31 GB | 19 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
Frequently asked questions
- What is Arete-27B?
- Arete-27B is Alibaba's Qwen3.8-27B retrained by Nous Research to give fewer refusals and fewer answers framed in the Chinese government's terms on politically sensitive topics. Nous describes this as reinforcing Western values. The weights are free under Apache 2.0, and the test results come from Nous itself.
- Can I use Arete-27B commercially?
- Yes. Arete-27B is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Arete-27B open source?
- Arete-27B is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Arete-27B on my own computer?
- Yes, on a good graphics card. The smallest variant, 27B, needs about 19 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Arete-27B released?
- Arete-27B was released by Nous Research on November 6, 2026.
- How long a context window does Arete-27B have?
- Arete-27B supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Arete-27B model card (Hugging Face) huggingface.co
- Nous Research: Introducing Arete-27B, Qwen3.8-27B post-trained to reinforce Western values nousresearch.com
- Propaganda_Eval, the 803-prompt evaluation (Hugging Face dataset) huggingface.co
- arete-data-pairs, the training data (Hugging Face dataset) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Inkling Thinking Machines Lab | 276B – 975B 41B active | 1M | Apache 2.0 (Permissive) | 188 GB Big-memory desktop | |
| Mistral Large 4 Mistral AI | 1.05T 52B active | 1M | Mistral Open-Weight (Conditional) | 671 GB GPU server | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone |