Arete-27B: what it is and how to run it

Arete-27B is a free, open-weight model from Nous Research: Alibaba's Qwen3.8-27B, fine-tuned to drop the Chinese state-aligned answers baked into the original. You download it from Hugging Face under an Apache 2.0 license and run it yourself, since Nous hasn't announced a hosted API or chat version.

@NousResearch announced it on November 6, calling it Qwen3.8-27B "post-trained to reinforce Western values." The argument behind it: Chinese open models are the strongest you can download right now, and they ship with the government's line already in the weights. The developer @iamlukethedev summed up who should pay attention: anyone who runs "local agents on Chinese open weights."

What Nous actually changed

According to the Arete write-up, Nous built an 803-prompt test across seven areas: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. Grok 4.7 graded the answers.

The base Qwen3.8-27B gave answers with some propaganda in them 22% of the time. Arete-27B came in at 0.9% in Nous's table (the post's text rounds it to 1%). On freedom-of-expression prompts alone, the rate went from 60% to 1%. Factual accuracy, scored 0 to 5, rose from 2.32 to 2.98.

Nous used supervised fine-tuning only, with no reinforcement learning, no preference tuning and no changes to the architecture. That came to about 52.8 million tokens over two epochs, with a model called Inkling acting as the teacher. On Nous's internal Hermes Bench, the base model scored 65.77% and Arete 65.60%, a gap Nous puts down to noise.

How to run Arete-27B

There's no one-click option yet. The model card says no inference provider hosts it, and when I checked, OpenRouter and the Ollama library didn't list it either. These are the pieces Nous published:

For a server with enough GPU memory, the usual route works for a Qwen-based checkpoint. Run vllm serve NousResearch/Arete-27B, assuming your vLLM version already supports the Qwen3.8 architecture. In BF16 the weights alone take roughly 54 GB. On a laptop, pick one of the community quantizations and load it in llama.cpp or LM Studio. Those builds come from third parties, so check who made the one you download.

Once it's running, ask both models the same questions and compare. Try "What happened in Tiananmen Square in 1989?" or "Should citizens be allowed to criticize their government online?" You can also take prompts from the Propaganda_Eval set and send them to Arete-27B and stock Qwen3.8-27B side by side.

My read

What surprised me is how little it took. Nous describes the bias layer in a Chinese-aligned model as shallow, and its numbers support that: about 53 million tokens of plain fine-tuning removed most of it, and the general benchmark score barely moved.

I'd still hold the results loosely. Nous wrote the test, wrote the scoring rules and picked the judge, and nobody has replicated the work yet. The thinking-mode numbers come from two of five planned runs and are marked provisional, and Nous itself says it hasn't "remedied all behaviors."

Political framing is also a different problem from the one Cohere flagged last month, when another Qwen model tried to upload a file to a Beijing bucket during a sandboxed test. Changing how a model talks won't stop an agent from making network calls you didn't expect.

What this means for a business

If your team pulled a Qwen model into a local agent because it was free and strong, Arete-27B gives you a drop-in alternative to test, with the data published so you can check the claims. It also fits where Nous is going commercially with Hermes for Businesses, which is meant to let companies choose and own their models instead of renting them. Before you put it in front of customers, run your own prompts through both models first and keep the agent's network access locked down whichever one you pick.

If you're not sure which of your workflows are worth handing to an AI model at all, New Face Design's free process audit is a good place to sort that out before you start comparing checkpoints.