Arete-27B: Nous retunes Qwen3.8-27B, Apache-2.0, flagged propaganda 22% to 1% on its own test
Nous Research released a supervised fine-tune of Qwen3.8-27B aimed at politically sensitive prompts; all results come from Nous's own judge and prompts, with no independent test yet.
· 7 min read
Written by AI from the sources linked below. How we use AI
On this page
Nous Research has released Arete-27B, a version of Alibaba's Qwen3.8-27B that it post-trained to change how the model answers questions on politically sensitive topics. The weights are on Hugging Face under the Apache-2.0 licence, and Nous says that on its own 803-prompt test the share of answers it flagged as propaganda fell from 22% to 1%. Every result below comes from Nous and its own evaluation; we found no independent test yet, so read the numbers as the lab's claims, not settled facts.
Key points
- Arete-27B is a 27-billion-parameter fine-tune of Qwen/Qwen3.8-27B, trained with supervised fine-tuning only (no reinforcement learning), and published as NousResearch/Arete-27B on Hugging Face under Apache-2.0.
- On Nous's Propaganda_Eval (803 prompts, seven topic categories, scored by an LLM judge), the base model's fact score of 2.32 out of 5 rose to 2.98, and answers flagged for propaganda fell from 22% to 1% (thinking off, temperature 0).
- General ability looks unchanged on Nous's internal Hermes Bench: 65.60% for Arete against 65.77% for the base model, a gap the authors call within noise.
- Nous says the training took 1,202 steps over two epochs on text-only data, with a larger open model, Inkling, as the teacher.
- The model card says plainly that not every behaviour tied to the original viewpoint was removed, and the judge, prompts and settings are all Nous's own.
What happened
In a November 2026 post, Nous Research introduced Arete-27B. The company's argument is that the strongest open-weight models are often Chinese, and that those models can carry state-aligned behaviour. In its tests, Qwen3.8-27B produced refusals, deflections and answers framed the way Chinese state media frames them on sensitive subjects. Nous chose that model as its starting point because, it says, it was the most capable local model that showed these behaviours.
The fix is a modest fine-tune. Nous generated about 20,000 prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. For each prompt it collected a response from a teacher model, Inkling, and from the base model itself. It then trained all of the language model's weights on the teacher's answers. The company says it picked an open teacher on purpose, to avoid terms-of-use questions around distilling from a closed model. It also says it considered NVIDIA's Nemotron 3 Ultra as a teacher and rejected it after finding that model had its own China-related bias, which the authors suspect came from contamination via Chinese open models.
The result is published as weights (BF16, Safetensors) plus the evaluation set, the judge prompt, the grading scripts and the training data pairs, all linked from the Nous post. The model card lists the intended uses as general assistance, agentic work and research into value alignment.
Background
Qwen3.8-27B is the base. Its own model card lists a 27-billion-parameter model with a vision encoder, 64 layers, and a hybrid design that alternates three Gated DeltaNet layers with one Gated Attention layer. It handles 262,144 tokens of context natively and up to 1,000,000 with YaRN scaling, and it is released under Apache-2.0, which is why a third party could retrain it and republish the result.
The idea that language models absorb a point of view from their training data and tuning is not new. Nous's claim is narrower and testable: that in a Chinese-aligned model the relevant behaviour sits in a thin layer, so a small amount of supervised training can shift it without damaging anything else. The authors' own summary is that "the bias layer in a Chinese-aligned model is shallow."
The numbers
Nous built Propaganda_Eval for this release. It has 803 prompts. Each prompt carries two to four required facts taken from a dated source article (English-language statutes, news and research reports, with no source supplying more than 13 prompts). A judge model, Grok 4.7, marks each fact as stated, missed or opposed, and flags propaganda only when it can quote a supporting phrase. Nous says the judge is not told which model wrote an answer and that it tuned the judge on a hand-graded pilot of 25 prompts, revising the rubric three times. The "fact mean" runs from 0 to 5, and "any propaganda" is the share of answers scoring 1 or 2; unanswered prompts count as 0.
Several details matter when reading these. Nous ran five runs per model with a 32,000-token cap, and most rows use temperature 0.6, while the two Arete and base rows quoted in the headline use temperature 0. At temperature 0.6 the same two models score 2.88 and 2.24. The "thinking on" Arete row is marked provisional because it covers only two of five runs. The biggest change came in the freedom-of-expression category, where the fact score rose from 0.89 to 2.44 and flagged answers fell from 60% to 1%. Answers scoring zero dropped from 359 to 216 of the 803 prompts, and Nous says the improvement held across all seven categories.
On general capability, Nous reports Hermes Bench, an internal benchmark, at 65.60% for Arete and 65.77% for the base model. On Terminal-Bench-2.1 it reports 70% against 67%, but it ran both models on its own hardware and says it does not claim that gain is real, attributing differences from official numbers to inference and harness setup.
The training recipe is plain. The optimizer is AdamW with weight decay 0.1 and gradient clipping at 1.0. The learning rate warms up from 1e-6 to 1e-5 over 120 steps and then decays on a cosine schedule to 1e-6 at step 1,202. Batch size is 48, training runs two epochs of 601 steps each in bf16 with fp32 master weights, with tensor parallelism of 4. The vision encoder was frozen because the data is text-only, and the multi-token-prediction head was switched off during training.
What it means
For people who run open models, the practical point is that this is a drop-in replacement candidate rather than a new architecture. It is the same size and family as Qwen3.8-27B, so anything that serves the base model should in principle serve Arete. Qwen's card documents vLLM (vllm serve "Qwen/Qwen3.8-27B"), SGLang, Hugging Face Transformers and Docker Model Runner, and says quantised builds work with llama.cpp, Ollama and LM Studio. The Arete card, by contrast, includes no run instructions and notes that no hosted inference provider currently serves it; the Hugging Face page lists 32 quantised models derived from it. We have not run the model, so treat compatibility with those tools as an expectation, not a tested fact. As a rough size guide, 27 billion parameters in BF16 is about 54 GB of weights before any memory for context, so most people will want a quantised build.
For researchers, the release is as interesting for its evaluation as for its weights. Publishing the prompts, the judge prompt and the grading scripts lets others rerun or attack the test. That matters because the headline measurement depends on an LLM judge whose rubric was tuned on 25 prompts, and because the judge, Grok 4.7, is a third-party model.
There is also a values question the numbers cannot settle. Nous is not neutral about the outcome: its stated aim is to shift how the model treats topics such as democracy and free expression, and the test is built from English-language sources. Someone who wants a model that mirrors a different set of views, or one that stays out of such topics, will read the same results differently. What the data can support is narrower: on this test, this fine-tune answers the sensitive questions more often and with fewer flagged framings.
What we don't know yet
- Independent results. We found no outside evaluation of Arete-27B. All scores, the judge and the prompts come from Nous.
- The training set size. The post gives about 52.8 million tokens and roughly 50,000 samples at 64,000 context in one place and 28,886 training examples in an interactive caption, and does not reconcile them.
- How much was left. The authors say they did not remedy all behaviours associated with the original viewpoint, and they do not say which remain.
- Real-world behaviour. An 803-prompt test with English sources and an LLM judge says little about other languages, long conversations, or adversarial prompting. The training data was text-only, so image behaviour was not trained or measured here.
- Comparability. Temperature and thinking settings differ between rows, so the cross-model ranking is looser than the table suggests.
What to watch next
Look for independent reruns of Propaganda_Eval with a different judge model, results on standard safety and knowledge benchmarks to confirm nothing else moved, and community quantisations with published speed and memory figures. Also watch whether other labs adopt the same recipe for other base models.
Sources
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
Inkling 2 Small: Thinking Machines' 594B model scores 31 on AA index, Apache 2.0 weights promised next week
Thinking Machines Lab's second-generation model is on Tinker now with closed weights; Artificial Analysis says it would top open models outside China if weights ship.
audiotale.net team6 min read
Nvidia releases Nemotron 3.5 Super VL and Bolt open models for enterprise and local agents
Super VL has 120B parameters (12B active); Bolt 74B-A9B fits in 24 GB with 2-bit weights. Both ship under OpenMDW-1.1, and the speed claims are Nvidia's own.
audiotale.net team5 min read
DeepSeek releases V4.1-Flash-1017 weights under MIT licence, claiming large gains on agent benchmarks
The October update to DeepSeek's 552-billion-parameter mixture-of-experts model keeps its price and million-token context, but every benchmark gain so far is DeepSeek's own.
audiotale.net team10 min read