What is Arete-27B?
Nous Research released Arete-27B on 6 November 2026, describing it as a post-trained version of Alibaba's Qwen3.8-27B. The company says the new model keeps the base model's abilities while removing answers that adopt Chinese state framing. The weights, the 803-prompt evaluation set and the training data are all public.
Arete-27B is a 27B dense model under the Apache 2.0 licence, the same licence Qwen3.8-27B carries. Qwen's model card describes the base as a native vision-language system. Its context window is 262,144 tokens, extensible to one million. Nous trained only the language weights, left the vision encoder frozen and used text-only data.
Post-training is the stage after a model's initial, expensive training run. It can adjust behaviour with far less compute than pretraining, which is why a smaller lab can reshape a large open model. Arete is a demonstration of that leverage more than a new architecture.
The name comes from the ancient Greek word arete, which Nous translates as excellence or virtue. The company also notes that an arete is a sharp ridge between two glacial valleys. That image fits the pitch: a model positioned between two political vocabularies.
The headline result is a change in measured framing. On Nous's evaluation, answers containing what the company calls propaganda fell from 22% to 0.9%. The fact score rose from 2.32 to 2.98 out of 5, and the improvement held across all seven categories.
Those numbers come from the company that built and tested the model. They are not an independent audit, and Nous presents several caveats alongside them. The most important is that the evaluation measures framing, not overall truthfulness.
Who is Nous Research, and why Qwen?
Nous Research is an open-weight AI lab best known for its Hermes models and agent software. It has built a following by releasing models that developers can download and modify, rather than only renting them through an API. Crypto Briefing notes that the company has consistently operated in the open-weight space.
Qwen3.8-27B was an obvious base. Alibaba's Qwen family is widely used for local deployment. The 27B release is one of the stronger dense models at that size. Hugging Face listed about 7.8 million downloads and 19,000 likes for it on 8 November.
The base model is also tuned for agentic work. Alibaba's model card highlights coding, long-horizon tasks and flexible thinking control. It also says reasoning context carries across messages, which Nous is trying to preserve.
Nous says it chose Qwen3.8-27B because it was the most capable local model it tested. It also showed the behaviour the team wanted to change, including refusals, deflections and answers that echoed Chinese Communist Party positions.
The concern is not new. The Decoder notes that NewsGuard tested five leading models, including Qwen, in 2025. NewsGuard found they repeated or failed to correct pro-Chinese disinformation in 60% of cases. The outlet also cites CrowdStrike research on DeepSeek producing less secure code when queries mention Taiwan, Tibet or Falun Gong.
Nous frames its work as a response to that behaviour. Critics may see a political project instead, because the model is deliberately trained toward one set of values rather than toward neutrality. Nous does not hide that goal; it argues that users deserve a choice of viewpoints.
How did Nous remove the state-aligned answers?
The method is supervised fine-tuning, not reinforcement learning. Nous generated 20,000 prompts across seven value categories, filtered duplicates and removed any prompt that overlapped the evaluation set. It then collected answers from a teacher model and from the starting Qwen model.
The teacher was Inkling, an open model from Thinking Machines. Nous says it picked an open teacher to avoid licence terms that restrict distillation. It also says it considered Nemotron 3 Ultra but found evidence of a checkered record on China-related topics.
That Nemotron claim is a suspicion, not a finding. Nous writes that it does not know the cause and theorises that bias carried over from training on Chinese open models. The Decoder adds that Thinking Machines used Chinese models, including Kimi K2.5, in Inkling's synthetic data.
Training used about 52.8 million tokens at a 64,000-token context, without sequence packing. The run covered two epochs, 1,202 steps and a batch size of 48. Nous changed no architecture, added no preference optimisation and trained every language-model weight.
Nous says every batch of training data moved through reviewed pull requests, the same way the team ships code. That is a process claim, not a technical guarantee. Still, it makes the released dataset easier to audit.
Loss fell quickly during a 120-step warmup and levelled near 1.0 by the end of the first epoch. It stepped down to about 0.8 when the second epoch began. It stayed flat because the model was seeing the same examples again. Gradient norm started near 32 and settled around 1.4.
The released evaluation is MIT licensed and includes the prompts, required facts, judge prompt and scoring scripts. That matters more than the training recipe, because it lets outsiders repeat the test instead of trusting the summary.
Do the 22% and 0.9% numbers hold up?
The evaluation has 803 prompts across seven value categories. They cover free expression, state ideology, state versus free market, education, data privacy, democracy and the environment. Each prompt carries two to four required facts from a dated source. No single source supplies more than 13 prompts.
An LLM judge, Grok 4.7, graded each answer without being told which model produced it. Nous built the rubric on a 25-prompt pilot that humans graded by hand. The team revised the rubric three times where the judge and the human grades disagreed.
The comparison table is the most revealing part. Arete-27B scores 3.07 with thinking on, but that figure is provisional and based on two of five runs. With thinking off it scores 2.98, against 2.32 for Qwen3.8-27B.
Not every model with a higher fact score is more state-aligned. GLM 5.3 reaches 3.10 but has a 15% propaganda rate. Mistral Large 4 reaches 3.65 with 0.3%, and Inkling, the teacher, reaches 3.96 with 0.3%.
So the fine-tune appears to change framing more reliably than it improves factual accuracy. Arete also leaves 216 of 803 answers at a fact score of zero, down from 359. The model is less state-aligned, not necessarily more truthful.
There are methodological wrinkles. Arete and Qwen ran at temperature 0, while several comparison models ran at 0.6. Nous discloses this and reports a second Qwen comparison at 0.6, where Arete scores 2.88 and Qwen 2.24.
The propaganda score is not a general bias detector. An answer scores above zero only when it states the official line or endorses it. The judge must quote the exact phrase. Quoting the line to identify it does not count, which narrows the metric to adoption rather than mention.
The biggest change appeared in freedom of expression. There the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%. That category may be the clearest evidence that the training changed behaviour rather than noise.
The bigger caveat is authorship. Nous wrote the prompts, chose the judge and ran the tests. Independent replication has not been published yet, so the numbers should be read as a vendor's results rather than a settled measurement.
Why does an open-weight fine-tune matter?
The release lands in the middle of a policy fight over Chinese models. In October, the US Department of Homeland Security made public an order banning DeepSeek-related technology from federal civilian agency IT systems. The order was issued on 24 July and excludes the Department of Defense.
The order also covers third-party products and services built with DeepSeek technology, according to CNBC. It names DeepSeek, not Qwen. It governs procurement and federal systems, not the download of open weights.
Congress had already restricted DeepSeek in intelligence community systems in December 2025, CNBC reports. The newer order covers civilian agencies such as Treasury, State and Justice.
That gap matters. Procurement rules can keep a model out of a government laptop, but they cannot recall weights that are already public. Fine-tunes can move a model toward or away from a government's preferred framing, and the licence allows it.
China's foreign ministry pushed back on the DeepSeek order. A spokesperson told Reuters that “China consistently opposes the over-generalization of national security.” The statement also rejected the politicization of science, technology and trade issues, CNBC reported.
For companies, Nous positions Arete as Qwen's agentic and coding ability with a different political default. It is aimed at general assistance, agentic work and research on how post-training changes behaviour. No hosted inference provider served the model when MarkTechPost checked.
Running it locally is feasible but not trivial. MarkTechPost reports BF16 weights at 53.81 GB and a community Q4_K_M quantisation at 17.20 GB. That quantisation is sized for a 24 GB consumer GPU.
The most consequential claim is about durability. Nous argues that “the values in any open model are only ever one fine-tune away.” If that is right, alignment is a property of the checkpoint, not of the base model or its country of origin.
What are the limitations and open questions?
“Western values” is a broad label for a narrow test. The evaluation measures whether answers adopt Chinese state framing on seven topics. It does not measure whether a model favours American, European or other political positions.
That distinction matters because Nous is not aiming for neutrality. The Decoder notes that Hermes 3, released in 2024, was designed to adapt to a worldview set in the system prompt. Arete instead ships with a preferred worldview baked into the weights.
Nous calls the result evidence that “the bias layer in a Chinese-aligned model is shallow.” The phrase is doing a lot of work. The test covers seven English-language categories, not the full range of political or cultural behaviour a model might show.
The training data is text-only and English. The base model accepts images and video, but Nous froze the vision encoder and did not evaluate multimodal prompts. Behaviour on Chinese-language questions is also untested in the published evaluation.
The Nemotron contamination story remains unproven, and Nous does not name the evidence. The teacher's own lineage is another open question, since The Decoder reports that Inkling's synthetic data involved Chinese models.
Finally, a 0.9% rate is not zero. A small number of answers still met Nous's definition of propaganda, and that definition depends on quoted phrases and a judge model. Small changes in the rubric could move the score.
The licence also cuts both ways. Apache 2.0 lets anyone distribute a modified model, including one trained toward a different ideology. Nous presents that as optionality; a regulator may see it as an unenforceable safeguard.
Nous does not report a safety evaluation. The published benchmarks focus on general capability and agentic tasks; neither is a safety test. A fine-tune that changes political framing could, in principle, change other behaviours too.
What should you watch next?
The first test is replication. The evaluation is open and MIT licensed, so outside researchers can run it on Arete, on Qwen3.8-27B and on other models. A result that holds under a different judge and prompt set would carry far more weight.
The second is distribution. Community quantisations appeared within a day of release, but hosted inference had not arrived when MarkTechPost published. Enterprise adoption will depend on serving cost, latency and support, not just framing.
The third is emulation. If a 52.8-million-token fine-tune can shift a model's framing this much, other labs may try the same on GLM, DeepSeek or Kimi. That would turn value alignment into a competitive feature rather than a fixed property.
The fourth is policy. Washington is restricting DeepSeek in federal systems while open weights circulate freely. Arete shows how quickly a downloaded model can be repurposed, which is exactly the scenario procurement rules struggle to control.
The fifth is Alibaba's response. Qwen3.8-27B remains available under Apache 2.0, so Nous needed no permission. A future release could add licence terms, change the base behaviour or make derivatives harder to distribute.
One near-term update is already promised. Nous says it will update the thinking-on score when the remaining runs finish. That will turn a provisional 3.07 into a settled figure either way.
For now, the honest summary is narrow. Nous has shown that one measured behaviour can be reduced on one model with a modest fine-tune. Whether that generalises to other behaviours, languages and auditors is still open.