{"slug": "arete-27b-vs-qwen3-8-27b-propaganda-down-to-0-9", "title": "Arete 27B vs Qwen3.8-27B: Propaganda Down to 0.9%", "summary": "Nous Research released Arete 27B, a supervised fine-tune of Qwen3.8-27B trained on roughly 52.8 million tokens to reinforce Western values in a Chinese-trained model. On Nous's own values evaluation, the share of answers endorsing or substituting an official party line fell from 22% to 0.9%, while the mean fact score rose from 2.32 to 2.98 out of 5, with general capability essentially unchanged (Hermes Bench 65.60% versus 65.77% for the base). The Apache 2.0 weights, training pairs and evaluation set are published on Hugging Face, though independent replication has not yet appeared.", "body_md": "Arete 27B beats its base model, Qwen3.8-27B, on Nous Research's own values evaluation. The share of answers that endorse or substitute an official party line fell from 22% to 0.9%, and the mean fact score rose from 2.32 to 2.98 out of 5 ([Nous Research](https://nousresearch.com/arete)). Capability looks unchanged and the weights are Apache 2.0, so the verdict: if you already run Qwen3.8-27B locally and care about politically sensitive topics, switch to Arete 27B. If you do not, the base model remains the safer default.\n\nArete 27B is Nous Research's post-train of Qwen3.8-27B, described as an attempt to reinforce Western values on a Chinese-trained model. The name comes from the Greek word for excellence or virtue. The recipe is deliberately narrow: supervised fine-tuning only, with no reinforcement learning, no preference optimisation and no architectural change ([Nous Research](https://nousresearch.com/arete)).\n\nThe training run used about 52.8 million tokens over 2 epochs and 1,202 steps at 64k context. All language-model weights were trained; the vision encoder stayed frozen because the data was text-only. The teacher was Inkling, which Nous picked as the best open model on its own evaluation. Weights are on [Hugging Face](https://huggingface.co/NousResearch/Arete-27B), alongside the training pairs and the evaluation set.\n\nRoughly 53 million tokens is small by modern standards ([Nous Research](https://nousresearch.com/arete)), which suggests the \"values layer\" of an open model can be rewritten cheaply once someone has a recipe.\n\nOn Nous's values eval, Arete 27B improves on the base model in all seven categories. The table uses the fact mean (0 to 5) and the share of answers containing any propaganda, averaged over 5 runs of 803 prompts ([Nous Research](https://nousresearch.com/arete)).\n\n| Model | Fact mean | Any propaganda | \n|---|---|---|\n| Inkling (teacher) | 3.96 | 0.3% | \n| Mistral Large 4 | 3.65 | 0.3% | \n| Nemotron 3 Ultra | 3.22 | 0.6% | \n| GLM 5.3 | 3.10 | 15% | \n| **Arete 27B (thinking off)** | **2.98** | **0.9%** | \n| DeepSeek V4.1 Flash | 2.39 | 16% | \n| Qwen3.8-27B (base) | 2.32 | 22% | \n\nThe largest single gain was in freedom of expression, where the fact score moved from 0.89 to 2.44 and the propaganda rate from 60% to 1%. Answers scoring zero fell from 359 to 216 of 803 ([Nous Research](https://nousresearch.com/arete)). At temperature 0.6 the gap holds, at 2.88 for Arete against 2.24 for the base.\n\nArete's propaganda rate matches the Western-trained models, but its fact score sits closer to GLM 5.3 than to Mistral Large 4: it stopped repeating the official line more reliably than it learned the full set of facts.\n\nNo evidence of that so far. On Hermes Bench, Nous's internal general benchmark, the base scores 65.77% and Arete 65.60%, a gap of under 0.2 percentage points that Nous calls noise. On Terminal-Bench 2.1 the base scored 67% and Arete 70%, but Nous explicitly declines to claim that gain because the harness configuration differed ([Nous Research](https://nousresearch.com/arete)).\n\nThe values eval was built, run and published by Nous, with Grok 4.7 as a blind judge and a rubric refined on a 25-prompt pilot. Two of the three capability numbers come from an internal benchmark. Independent replication has not appeared yet, so treat the results as a credible first report rather than settled fact — third-party coverage agrees on the numbers ([MarkTechPost](https://www.marktechpost.com/2026/11/07/nous-research-arete-27b-qwen3-8-27b-propaganda-fine-tune/)). Also note that \"Western values\" is a framing choice; the eval measures factual accuracy against dated English-language sources, not a universal standard of truth.\n\nThe fastest route is Ollama, using the community GGUF quantisations by bartowski, which were built with llama.cpp b12174 ([Hugging Face](https://huggingface.co/bartowski/NousResearch_Arete-27B-GGUF)):\n\n```\nollama run hf.co/bartowski/NousResearch_Arete-27B-GGUF:Q4_K_M\n```\n\nUpdate Ollama first, because the model's hybrid attention needs a recent build. For llama.cpp, this starts a server with a built-in web interface on `localhost:8080`:\n\n```\nllama-server -hf bartowski/NousResearch_Arete-27B-GGUF:Q4_K_M\n```\n\nFor multi-user serving, vLLM loads the full BF16 weights:\n\n```\nvllm serve NousResearch/Arete-27B --max-model-len 65536\n```\n\nThe BF16 file is about 54GB, so plan for a 64 to 80GB GPU or several smaller ones ([Hugging Face](https://huggingface.co/NousResearch/Arete-27B)). For the base model's own setup details, see our [Qwen3.8-27B local setup guide](https://dev.to/articles/qwen-3-8-27b-local-setup-guide-2026).\n\nQ4_K_M is the sensible default at 17.20GB on disk, and it needs about 19GB with an 8K context once you add KV cache and overhead ([bartowski GGUF](https://huggingface.co/bartowski/NousResearch_Arete-27B-GGUF)). A 24GB card handles 8K and 16K context at that quant, and 32K wants a 32GB card. Other file sizes from the same release:\n\nOn a Mac, 32GB or more of unified memory suits Q4 or Q5 ([bartowski GGUF](https://huggingface.co/bartowski/NousResearch_Arete-27B-GGUF)). Keep context at or under 64k. The base config advertises 262,144 positions ([Hugging Face](https://huggingface.co/NousResearch/Arete-27B)), but Nous trained at 64k, so longer windows are untested.\n\nArete 27B itself costs nothing to run beyond your own hardware and electricity, because the weights are Apache 2.0. No inference provider hosts it yet, and the Hugging Face card says so ([Hugging Face](https://huggingface.co/NousResearch/Arete-27B)).\n\nThe nearest hosted equivalent is the base model. The prices below are the OpenRouter listings observed on 23 October 2026 ([OpenRouter](https://openrouter.ai/qwen/qwen3.8-27b)), so check the live page before relying on them. Rupee figures use 93.83 INR per USD, the rate on 9 November 2026 ([Investing.com](https://au.investing.com/currencies/usd-inr)).\n\n| Option | Input per 1M | Output per 1M | \n|---|---|---|\n| Arete 27B, self-hosted | Free | Free | \n| Qwen3.8-27B, OpenRouter default | $0.20 (about Rs 19) | $2.50 (about Rs 235) | \n| Qwen3.8-27B, cheapest listed route | $0.15 (about Rs 14) | $1.88 (about Rs 176) | \n\nIf you generate far more output than input, which is typical for reasoning-heavy work, the output price dominates. For occasional use, the API route for the base model is cheaper than buying hardware, but it will not give you Arete's behaviour.\n\nWithin our 72 winnable keywords, 20 carry a difficulty score of zero, meaning no established competitor holds the result set (n=72, measured 2026-11-09, [Method](https://aitecharchive.com/pages/how-we-work)). That is exactly the slot Arete 27B lands in: it is a specialist variant, not a new frontier model. If raw capability per rupee is your question, compare architectures in our [Qwen3.8 Flash Next vs Qwen3.8-27B comparison](https://dev.to/articles/qwen-3-8-flash-next-vs-qwen-3-8-27b-125b-6b-active) and the wider [Qwen vs Claude vs Kimi](https://dev.to/articles/qwen-vs-claude-vs-kimi) review. For another fast-moving tier, see [GLM 5.3 Flash vs Qwen3.8 Flash Next vs DeepSeek V4 Flash](https://dev.to/articles/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-deepseek-v4-flash).\n\nArete also sits in the wider conversation about editing model behaviour after release: our piece on [Heretic 2.0 ARA vs abliteration](https://dev.to/articles/heretic-2-0-ara-vs-abliteration) covers the opposite direction, removing refusals; together they show how little effort it takes to reshape an open model. Nous itself says it addressed \"a potential root cause\", not cured every bias, and the remaining fact-score gap to Inkling supports that modesty.\n\n**Q: What is Arete 27B?**\n\n**A:** A supervised fine-tune of Alibaba's Qwen3.8-27B, released by Nous Research on 6 November 2026 under Apache 2.0, to reduce adoption of official party lines on politically sensitive topics.\n\n**Q: How do I run Arete 27B locally?**\n\n**A:** Run `ollama run hf.co/bartowski/NousResearch_Arete-27B-GGUF:Q4_K_M` on a recent Ollama build. The Q4_K_M file is 17.20GB and fits a 24GB GPU at 8K to 16K context ([bartowski GGUF](https://huggingface.co/bartowski/NousResearch_Arete-27B-GGUF)).\n\n**Q: How much does Arete 27B cost per 1M tokens?**\n\n**A:** Nothing if you self-host, since no provider hosts it yet. The base Qwen3.8-27B was listed on OpenRouter at $0.20 input and $2.50 output per 1M tokens, about Rs 19 and Rs 235, on 23 October 2026 ([OpenRouter](https://openrouter.ai/qwen/qwen3.8-27b)).\n\n**Q: Did Arete 27B become less capable than Qwen3.8-27B?**\n\n**A:** Not on the evidence Nous published. Hermes Bench scores were 65.77% for the base and 65.60% for Arete, and Terminal-Bench 2.1 read 67% against 70%, though Nous does not claim that gain ([Nous Research](https://nousresearch.com/arete)).\n\n**Q: Is Arete 27B available on OpenRouter?**\n\n**A:** Not as of 9 November 2026. The weights and community GGUF quantisations are on Hugging Face, so self-hosting is the only route.", "url": "https://wpnews.pro/news/arete-27b-vs-qwen3-8-27b-propaganda-down-to-0-9", "canonical_source": "https://dev.to/shaam_ai/arete-27b-vs-qwen38-27b-propaganda-down-to-09-399k", "published_at": "2026-11-10 04:44:16+00:00", "updated_at": "2026-11-10 04:50:43.257130+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-safety", "ai-ethics", "ai-tools"], "entities": ["Nous Research", "Qwen3.8-27B", "Arete 27B", "Inkling", "Hugging Face", "Mistral Large 4", "GLM 5.3", "DeepSeek V4.1 Flash"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/arete-27b-vs-qwen3-8-27b-propaganda-down-to-0-9", "markdown": "https://wpnews.pro/news/arete-27b-vs-qwen3-8-27b-propaganda-down-to-0-9.md", "text": "https://wpnews.pro/news/arete-27b-vs-qwen3-8-27b-propaganda-down-to-0-9.txt", "jsonld": "https://wpnews.pro/news/arete-27b-vs-qwen3-8-27b-propaganda-down-to-0-9.jsonld"}}