Arete-27B

Arete-27B

Answers politically sensitive questions with a 27B text model that Nous Research retrained from Qwen3.8-27B to shift its viewpoint. Made for assistant builders and value-alignment researchers.

Made by NousResearch on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Arete-27B is a 27B text generation model from Nous Research. It starts from Alibaba's Qwen3.8-27B and adds supervised fine-tuning. Nous Research says the goal was to reinforce western values, after seeing refusals, deflections and answers framed by the Chinese state on politically sensitive topics.

Who it suits: It suits people who want a model with this changed viewpoint, and people who study how extra training shifts a model's values. The maker names both uses.

Industry | Who uses it | How they use it | Example
AI research | Value-alignment researchers | Compare the tuned model with its starting model on political prompts | Checking fact recall across the seven value categories
Software | Developers building assistants | Use it as the model behind a general assistant or an agent | A chat assistant for users who want this viewpoint
Education | Instructors in civics or media literacy | Put answers from two models side by side as discussion material | Comparing replies on data privacy and surveillance

You can skip it if you want the original Qwen3.8-27B behavior or a model tuned for one narrow task.

Where to use it: It is a plain text generation model, so it fits any place you already run a model of this size.

• A chat assistant for general questions.
• An agent that plans and acts on tasks, which the maker names as an intended use.
• An evaluation script or notebook that compares it with Qwen3.8-27B.

Use case | Who | What they get back
General assistance | Users who want this viewpoint | Answers on political topics measured against the maker's fact rubric
Agentic work | Developers | A 27B model that kept its agentic benchmark results, with no meaningful drop reported by the maker
Alignment research | Researchers | A reference point for how fine-tuning on twenty thousand prompts shifts a model's values

How to set it up: The source shows no install command, so this page has no code block. Check the model page for loading code.

• A machine or server that can run a 27B parameter text model.
• Software you already use to run models of this size.

1. Download the Arete-27B weights from the model page on Hugging Face.
2. Load them in your usual model-running software.
3. Keep in mind that training used text only and a 64k context.
4. Point your chat tool or agent at the model.

How to use it: You use it like any text model. Send it questions or agent tasks, and read the answers with its deliberate viewpoint in mind.

Tool | What it does
Arete-27B | The tuned model, for chat and agent work
Propaganda_Eval | The maker's held-out test set of 803 prompts across the seven value categories
Blog post | The maker's write-up, which announces release of the training data

Example: You ask, "What rules govern state surveillance of citizens online in China?" The model is meant to state the relevant facts plainly instead of deflecting. The maker scores such answers from 0 to 5 on how many required facts they state.

For newcomers and experts: A model carries habits from its training, such as refusing a topic or repeating an official line. Arete-27B is a version where those habits were retrained. A good first try is a civics question, compared with the answer from the original Qwen3.8-27B.

If you know the field, this is supervised fine-tuning only, with no RL and no preference optimization. Training took two epochs, 1,202 steps in all, on roughly 52.8 million tokens. Sequences were 64k long and were not packed. All language-model weights were trained, the vision encoder was frozen, and the multi-token-prediction head was disabled. Inkling served as teacher, and its answers were collected next to the starting model's own. An LLM judge, Grok 4.7, scored answers without knowing which model wrote them.

Start from Qwen3.8-27B
Inkling answers twenty thousand prompts
Supervised fine-tuning, 1,202 steps
Judge scores 803 held-out prompts

Good to know:

• The viewpoint shift is deliberate. The maker says it addressed a likely root cause, not every behavior.
• The scores come from the maker's own evaluation and an LLM judge, so treat them as one lab's measurement.
• The maker reports no meaningful loss on the broad ability and agent tests it ran.
• The evaluation set is held out and is not the training data.
• The license is Apache 2.0.

Title: Fact score out of 5 on the maker's evaluation
Qwen3.8-27B: 2.32 points
Arete-27B: 2.98 points

Title: Share of answers containing propaganda
Qwen3.8-27B: 22%
Arete-27B: 0.9%

On Hugging Face, 10 November 2026

• Task: text generation
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,407
• Likes: 371
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 41
• Fine-tunes: 11
• Tags: safetensors, qwen3_5, nous-research, arete, supervised-fine-tuning, value-alignment, text-generation, conversational

Links

• Model: https://huggingface.co/NousResearch/Arete-27B
• Files: https://huggingface.co/NousResearch/Arete-27B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download NousResearch/Arete-27B

Then follow the usage in its model card. Nothing in this kit runs by itself.