Nous Research has released Arete-27B, a version of Alibaba’s open-weight Qwen3.8-27B model retrained to answer politically sensitive questions from a Western standpoint rather than Beijing’s. The US lab says the fine-tune cut answers adopting the Chinese state’s account from 22 percent to under one percent on its own 803-prompt test, while leaving the model’s general abilities intact. The weights, training data and test set are all public, so anyone can check the claim or run the model themselves.
What Nous actually did
Arete-27B starts from Qwen3.8-27B, Alibaba’s 27-billion-parameter open model, and changes nothing about its architecture. Nous applied supervised fine-tuning only — no reinforcement learning, no preference optimisation — training all of the model’s language weights on about 52.8 million tokens over two passes through the data. The vision encoder stayed frozen, since the training data was text only.
The training data came from 28,886 training examples built around seven value categories: free expression, state ideology, markets, education, privacy and surveillance, democracy, and the environment. Nous used Thinking Machines’ Inkling as the teacher model, chosen after tests showed it answered these topics cleanly. Every training prompt was checked against the test, and any overlap was removed before training began.
To measure the problem, Nous built Propaganda_Eval, an 803-prompt test where each question carries two to four required facts drawn from dated sources. An AI judge, Grok 4.7, graded answers blind, marking each fact stated, missed or contradicted, and quoting any phrase that adopted the official line. The base Qwen model scored 22 percent; Arete-27B scored 0.9 percent.
The headline numbers carry a qualifier: the test is Nous’s own, graded by an AI judge it chose, and no outside group has reproduced the results yet. General ability survived: on its internal Hermes Bench, Arete-27B scored 65.60 percent against the base model’s 65.77, a gap within noise, and 70 percent on Terminal-Bench-2.1 against the base model’s 67.
Freedom of expression moved most: fact score 0.89 to 2.44, propaganda rate 60 percent to 1 percent. Zero-scoring answers fell from 359 of 803 to 216. Nous is candid that the work is incomplete: it says it addressed a root cause, not every trace of the old viewpoint, and invites others to run its open test on any model.
Our Read
The result matters as proof, not product. A 27-billion-parameter model is small by frontier standards, yet fifty million tokens of supervision stripped out a viewpoint baked in by a top lab. Nous’s blunt conclusion: the bias layer in these models is shallow, applied late and thin, and the values in any open model are only one fine-tune away.
That cuts both ways. The technique that removes a state line can install one, and the open weights that let Nous audit Qwen let anyone retrain a model toward any agenda. Open models do not settle the values question; they move it from the training lab to whoever fine-tunes last. That is the real story, bigger than one 27-billion-parameter release.
The test measures a wider pattern. On those prompts, DeepSeek V4.1 Flash scored 16 percent and GLM 5.3 scored 15 percent, against 0.3 and 0.5 percent for Mistral Large 4 and Muse Glimmer 30B. Arete-27B sits with the Western group at 0.9 percent. The yardstick is Nous’s own, but the gap is the point: the slant is measurable and uneven.
Why It Matters
Governments are acting. Washington moved to keep DeepSeek off federal systems, and the debate covers every Chinese open model in the West. Arete-27B shows a third path: retrain the model instead of banning it or accepting its slant. If it holds up in independent testing, open models gain a way to keep the weights while choosing the values they carry.
Background
Qwen3.8-27B is Alibaba’s open-weight model, and it also powers Apple Intelligence in China under a deal announced on 31 October. Nous Research, known for its Hermes models, raised a reported $90 million Series B in October at a $1.5 billion valuation.
Source: Nous Research announcement · The Decoder report
0 Comments