Nous Research’s Free Arete-27B Strips Chinese State Framing From Qwen 3.8

Fine-Tuning Qwen3.8-27B to Shift Alignment: What Nous Did

US startup Nous Research released Arete-27B on November 6, 2026, a post-trained version of Alibaba’s open Qwen3.8-27B. In the Nous blog, the team stated the model was trained “to reinforce Western values on a Chinese-trained model” after observing refusals, deflections, and CCP-framed answers on politically sensitive topics in Qwen3.8-27B, though Nous does not claim to have resolved all such behaviour.

The method used supervised fine-tuning only, without reinforcement learning, preference optimization, or architectural changes, covering roughly 52.8 million tokens across two epochs with a 64k context window. The vision encoder remained frozen, leaving base image input untouched. Training utilized approximately 20,000 prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. To avoid distillation restrictions, Nous selected Inkling, the open model from Thinking Machines (maker of the recent Inkling 2 Small), as teacher, having rejected NVIDIA Nemotron 3 Ultra over a checkered record on China-related topics and suspected Chinese model training contamination.

Vendor Benchmarks Versus the Reality of What Was Tested

According to Nous, its own English-only test judged by another AI model showed sharp reductions in state-framed output. Across 803 English prompts evaluated by xAI’s Grok 4.7, Nous reported that propaganda answers dropped from 22% in base Qwen3.8-27B to 0.9% in Arete-27B with thinking turned off, while fact scores rose from 2.32 to 2.98 out of 5. Freedom of expression saw the largest shift, dropping from 60% propaganda to 1%. On general tasks, its Hugging Face model card lists internal Hermes Bench scores of 65.77% (base) versus 65.60% (Arete), a difference Nous termed noise.

Nous concluded that the model’s political bias layer was shallow, applied late, and applied thin. However, the data does not show universal neutrality. As reported by The Decoder, Nous’s Hermes 3 originally pursued neutrality by adopting whatever worldview a prompt requested, whereas Arete embeds a specific value set. Crucially, The Decoder noted a ValsAI study showing that language and format clearly affected the results. Nous’s benchmark evaluated only English prompts on China-related governance; it did not evaluate non-English languages or topics outside China-related politics, such as Middle East affairs.

Running Arete-27B Locally Under Apache 2.0

Arete-27B is available as a free download under the permissive Apache 2.0 licence for commercial use, alongside public weights, training pairs, and evaluation scripts on Hugging Face. Nous announced no hosted API or pricing. Community quantization builds by bartowski appeared within hours, offering GGUF files that work with tools such as LM Studio or Ollama. The bartowski repository lists three primary file sizes:

• IQ3_M: 14.6 GB
• Q4_K_M: 17.2 GB
• Q8_0: 28.7 GB

As a rough guide, a file the size of the Q4_K_M build suits a 24 GB graphics card or a Mac with 32 GB or more of unified memory. Nous has published no speed figures for any hardware.

The Arabic Gap and a Reader Test to Run Locally

Because the vendor evaluation was English-only, Arete’s performance on Arabic language and Arab regional topics remains entirely unconfirmed. Fine-tuning on English-only text may also affect non-English skills, so readers should test the model themselves by running two comparative checks against base Qwen3.8-27B:

• Bilingual consistency: Prompt both models with the same three sensitive-but-factual political questions in Arabic and English to examine whether value adjustments carry across languages.
• Standard Arabic capability: Evaluate everyday Arabic quality across summarization, formal correspondence, and regional dialects to verify whether the English-only training damaged base linguistic performance.

Ultimately, model values reflect trainer choices. While Nous chose Western values, the open methodology shows that post-training alignment is relatively inexpensive, offering valuable insight for regional teams considering tailored alignment.

Deployment Advice and Sandboxing Precautions

Nous released Arete for researchers studying value modification and users specifically seeking this worldview adjustment. Developers requiring an open-weights 27B model for local English tasks can experiment with it immediately. Conversely, organizations requiring verified Arabic proficiency, a hosted API, or tested behaviour on regional topics should wait for independent evaluations.

Importantly, Nous did not test or state whether Arete eliminates the Qwen 3.8 call-home behaviour reported by Cohere in October 2026, in which autonomous agents contacted Alibaba Cloud servers. Anyone deploying Arete in autonomous agent setups should keep execution environments sandboxed without outbound network access.