Nous Research just took Alibaba’s popular Qwen3.8-27B and retrained how it answers politically sensitive questions. The result is Arete-27B, an open-weights model under the Apache 2.0 licence that you can run on a home GPU.
Nous says the share of answers with state propaganda dropped from 22% to 0.9% on its own test, while coding and agent scores stayed the same. Community GGUF files appeared on Hugging Face within a day.
This guide covers what Arete-27B is, the full numbers (and their caveats), which GGUF fits your GPU, exact Ollama, LM Studio and llama.cpp steps, and how it compares with the base Qwen model and other 27-31B options.
- What: a supervised fine-tune of Qwen3.8-27B by Nous Research, published 6 November 2026 as NousResearch/Arete-27B.
- Goal: remove state-aligned framing on topics like free speech, democracy and surveillance, while keeping the model’s general skills.
- Headline result (vendor-reported): “any propaganda” rate 22% to 0.9%, fact score 2.32 to 2.98 out of 5, on Nous’s own 803-prompt test.
- Skills (vendor-reported): Hermes Bench 65.60% vs 65.77% for the base, Terminal-Bench-2.1 70% vs 67%.
- Run it: bartowski GGUF from 8.6 GB (IQ2_XXS) to 28.7 GB (Q8_0). IQ4_XS (15.2 GB) suits 16 GB cards, Q4_K_M (17.2 GB) suits 24 GB.
- Licence: Apache 2.0, same as the base model, so commercial use is allowed.
What Is Arete-27B?
Arete-27B is a post-trained version of Qwen3.8-27B, the 27-billion-parameter vision-language model Alibaba released in August 2026. It comes from Nous Research, the lab behind the Hermes model family.
The model card says the goal is to “reinforce western values on a Chinese-trained model”. Nous published the weights, the training data and its evaluation set on Hugging Face on 6 November 2026.
Why Nous built it
- The problem it targets: Nous says it saw refusals, deflections and state-framed answers from Qwen3.8-27B on sensitive political topics.
- Why it matters for open models: Qwen models are among the most used open weights, and many apps and agents are built on them.
- The bigger claim: Nous writes that “the bias layer in a Chinese-aligned model is shallow”, and that about fifty million tokens of supervision removes it “without disturbing anything underneath”.
The name comes from the Greek word for excellence or virtue. It is also the name of a sharp mountain ridge, which Nous mentions in its blog post.
Note: Arete-27B is not sold as an “uncensored” or abliterated model. Nous only describes changes to political and value topics. If you want models with refusals removed, see our best uncensored local LLMs guide instead.
How Nous Trained It
The method is simple on purpose. Nous used supervised fine-tuning (SFT) only: “No RL, no preference optimization, and no architectural modification.”
Data
- Prompts: about 20,000 generated prompts across seven value categories, checked for overlap with the test set.
- Categories: freedom of expression, state ideology, state vs free market, education, data privacy and surveillance, democracy, and environmental policy.
- Answers: written by a teacher model, Inkling from Thinking Machines Lab, which Nous calls “the best open model with a positive eval” on these topics.
- Size: about 52.8 million tokens of text-only data.
Nous says it avoided NVIDIA’s Nemotron 3 Ultra as a teacher because of what it calls irregularities on China-related topics. That is Nous’s own observation, not an independent audit.
Training settings
| Setting | Value |
|---|---|
| Epochs / steps | 2 epochs, 1,202 steps (601 per epoch) |
| Batch size | 48 |
| Learning rate | Warmup 1e-6 to 1e-5 over 120 steps, cosine decay to 1e-6 |
| Optimiser | AdamW, betas (0.9, 0.98), weight decay 0.1 |
| Precision | bf16 compute, fp32 master weights |
| Training context | 64k tokens |
| Frozen parts | Vision encoder frozen, multi-token prediction (MTP) head disabled |
Because the vision encoder was frozen and the data was text-only, image understanding should behave like the base model. Nous did not publish separate vision benchmarks.
The Results: Propaganda and Skill Scores
Nous tested on its own 803-prompt evaluation. Each prompt has 2-4 required facts from a dated source, and an LLM judge (Grok 4.7) scores each answer.
- Fact mean: 0 to 5, how many required facts the answer gets right.
- Any propaganda: share of answers the judge flags, and each flag needs a quoted phrase as evidence.
Before and after (vendor-reported, thinking off)
| Metric | Qwen3.8-27B (base) | Arete-27B |
|---|---|---|
| Fact mean (0-5) | 2.32 | 2.98 |
| Any propaganda | 22% | 0.9% |
| Answers scoring 0 | 359 of 803 | 216 of 803 |
| Freedom of expression: fact score | 0.89 | 2.44 |
| Freedom of expression: propaganda rate | 60% | 1% |
| Fact mean at temperature 0.6 | 2.24 | 2.88 |
With thinking turned on, Nous reports a provisional fact mean of 3.07 and 0.6% propaganda, based on 2 of 5 runs.
How it ranks against other open models
| Model | Fact mean | Any propaganda |
|---|---|---|
| Inkling (teacher) | 3.96 | 0.3% |
| Mistral Large 4 | 3.65 | 0.3% |
| Nemotron 3 Ultra (thinking off) | 3.22 | 0.6% |
| GLM 5.3 (thinking off) | 3.10 | 15% |
| Muse Glimmer 30B | 3.02 | 0.5% |
| Arete-27B (thinking off) | 2.98 | 0.9% |
| Gemma 4 31B | 2.58 | 0.5% |
| DeepSeek V4.1 Flash (thinking off) | 2.39 | 16% |
| Qwen3.8-27B (thinking off) | 2.32 | 22% |
All rows are from the Nous model card. Arete and the base were scored at temperature 0, the other models at 0.6, so the ranking across models is looser than the before/after comparison.
Did it lose any skills?
| Benchmark | Qwen3.8-27B | Arete-27B |
|---|---|---|
| Hermes Bench (Nous internal) | 65.77% | 65.60% |
| Terminal-Bench-2.1 | 67% | 70% |
Nous calls the Hermes Bench gap “noise” and says it does “not take credit” for the Terminal-Bench gain. In plain words: general and agent skills look unchanged.
How Much to Trust These Numbers
Nous is open about the limits, and they matter. Treat the 0.9% figure as Nous’s measured result, not an independent verdict.
- Same designer for test and training: Nous wrote the categories, rubric and prompts, and trained toward the teacher’s answers on those categories.
- LLM judge: Grok 4.7 did the grading. Nous checked the rubric against only a 25-prompt pilot it graded by hand, revising it three times.
- A stated viewpoint: “Western values” is Nous’s own framing, not a neutral standard. Different readers will judge the target answers differently.
- Partial fix: Nous says it addressed “a potential root cause” and did not remedy every behaviour tied to the original viewpoint.
- Own harness: the skill scores ran on Nous’s hardware with its own setup.
Warning: As of 8 November 2026 there is no independent re-run of the evaluation. The dataset and grading scripts are public, so expect community checks soon.
Arete-27B Specs at a Glance
| Spec | Arete-27B |
|---|---|
| Developer | Nous Research |
| Base model | Qwen/Qwen3.8-27B (Alibaba, August 2026) |
| Parameters | 27B (Hugging Face metadata shows 28B including vision parts) |
| Architecture | Hybrid: Gated DeltaNet linear attention + gated attention layers, 64 layers |
| Inputs | Text, images and video (vision encoder from the base, frozen) |
| Context | Base supports 262,144 tokens natively; Arete was trained at 64k |
| Thinking | On by default, can be turned off per request |
| Licence | Apache 2.0 |
| Release | 6 November 2026 |
| Hosted API | None at launch (no inference provider lists it) |
Nous does not state a maximum inference context for Arete. Since training used 64k sequences, staying at or under 64k is the safe choice for long chats.
Which GGUF Fits Your GPU
Within hours of launch, several quantised versions went up. The most complete set is bartowski/NousResearch_Arete-27B-GGUF, made with llama.cpp release b12174. It also ships the vision projector files.
| Quant | File size | Good for |
|---|---|---|
| Q8_0 | 28.67 GB | 32 GB cards or 2x 16 GB, near-lossless |
| Q6_K | 23.62 GB | 24 GB with short context, or 2x 16 GB |
| Q5_K_M | 20.68 GB | 24 GB cards with room for context |
| Q4_K_M | 17.20 GB | 24 GB cards, the default pick |
| IQ4_XS | 15.24 GB | 16 GB cards (tight, small context) |
| Q3_K_L | 13.88 GB | 16 GB cards with more context room |
| IQ3_XXS | 12.08 GB | 12 GB cards with partial offload |
| IQ2_M | 10.28 GB | Last resort for 12 GB |
Sizes are from the bartowski repo. You also need room for the KV cache, so leave 1-3 GB free on top of the file size (our estimate; it grows with context length).
- 16 GB (RTX 5080, 5070 Ti, 5060 Ti 16 GB): IQ4_XS for quality, Q3_K_L if you want 16k+ context on the GPU.
- 24 GB (RTX 4090, 3090): Q4_K_M or Q5_K_M.
- 2x 16 GB: Q6_K split across both cards with llama.cpp.
- Mac with 32 GB+ unified memory: Q4_K_M or Q5_K_M; bartowski notes Q4_1 has better tokens per watt on Apple silicon.
Tip: Arete uses the same architecture as Qwen3.8-27B, so every tip in our Qwen3.8-27B on a 16 GB GPU guide (offload, context, KV cache) applies here too.
How to Run Arete-27B Locally
Pick whichever tool you already use. If you are unsure, our vLLM vs llama.cpp vs Ollama comparison explains the trade-offs.
Option 1: Ollama (easiest)
- Update Ollama. Qwen3.8’s hybrid attention needs a recent build, so install the latest version from ollama.com.
- Pull and run the GGUF straight from Hugging Face. Change the tag to the quant you picked.
ollama run hf.co/bartowski/NousResearch_Arete-27B-GGUF:Q4_K_M - For a 16 GB card, use the smaller quant instead:
ollama run hf.co/bartowski/NousResearch_Arete-27B-GGUF:IQ4_XS - Chat in the terminal, or connect a front end such as Open WebUI to
http://localhost:11434.
Option 2: LM Studio
- Open the Discover tab and search for
Arete-27B. - Choose the bartowski upload and pick a quant. LM Studio marks which ones fit your VRAM.
- Load it in the Chat tab. Set context to 16k-32k to start, and raise it only if VRAM allows.
Option 3: llama.cpp (most control)
- Get a recent build. Use release b12174 or newer, the version the quants were made with.
- Download the file with the Hugging Face CLI:
pip install -U "huggingface_hub[cli]" hf download bartowski/NousResearch_Arete-27B-GGUF --include "NousResearch_Arete-27B-Q4_K_M.gguf" --local-dir ./ - Start the server. The
-hfflag also downloads the vision projector automatically:llama-server -hf bartowski/NousResearch_Arete-27B-GGUF:Q4_K_M -ngl 99 -c 32768 - For a local file plus images, pass the projector yourself:
llama-server -m NousResearch_Arete-27B-Q4_K_M.gguf --mmproj mmproj-NousResearch_Arete-27B-f16.gguf -ngl 99 -c 32768 - On 2x 16 GB GPUs, add
--split-mode layerso llama.cpp spreads the layers over both cards. - Open
http://localhost:8080for the built-in chat, or point any OpenAI-compatible app athttp://localhost:8080/v1.
Option 4: vLLM or SGLang (servers)
The full bf16 weights are about 54 GB, so this route needs a 64-80 GB GPU or several cards. The Qwen3.8-27B model card links official vLLM and SGLang recipes, and the same commands work with the Arete repo name.
vllm serve NousResearch/Arete-27B --max-model-len 65536Best Settings and Thinking Mode
Nous did not publish recommended sampling settings. It only lists its test setup: thinking off, temperature 0, 32k token cap. The base model’s official settings are the best starting point.
| Setting | Thinking on | Thinking off |
|---|---|---|
| Temperature | 1.0 | 0.7 |
| top_p | 0.95 | 0.80 |
| top_k | 20 | 20 |
| min_p | 0.0 | 0.0 |
| presence_penalty | 0.0 | 1.5 |
Source: Qwen3.8-27B model card. Thinking is on by default; the chat template opens each reply with a <think> block.
How to turn thinking off
- OpenAI-style API (vLLM, SGLang): send
"chat_template_kwargs": {"enable_thinking": false}in the request body. - llama.cpp server: start it with
--chat-template-kwargs '{"enable_thinking":false}'. - Reasoning depth: the base model accepts
reasoning_effortvaluesxhigh(default),mediumorlow.
Tip: To match Nous’s published numbers when you test, use thinking off and temperature 0.
Arete-27B vs Qwen3.8-27B and Other Options
| Model | Size | Licence | Propaganda rate (Nous test) | Best for |
|---|---|---|---|---|
| Arete-27B | 27B | Apache 2.0 | 0.9% | Qwen-level skills with fewer state-framed political answers |
| Qwen3.8-27B | 27B | Apache 2.0 | 22% | Official support, newest updates, same skills |
| Gemma 4 31B | 31B | See Google model page | 0.5% | Low propaganda rate, but lower fact score (2.58) |
| Muse Glimmer 30B | 30B | See Meta model page | 0.5% | Slightly higher fact score (3.02), in the Ollama library |
| Abliterated / Heretic Qwen | 27B | Apache 2.0 | Not tested | Removing refusals in general, not political framing |
Propaganda rates are from the Arete model card. For licence details on Gemma and Muse, check each model page before commercial use.
Who should use which
- Pick Arete-27B if you build chatbots, research tools or classroom apps where answers on history, rights or policy matter.
- Stay on Qwen3.8-27B for pure coding or agent work; Nous’s own numbers show no skill gain to switch for.
- Pick an abliterated model only if refusals are your problem. Arete is a different kind of change.
For more 16-24 GB picks by task, see our best Ollama models by VRAM list.
How to Check the Claims Yourself
Nous published everything needed to re-run its test. This is the fastest way to see if Arete behaves as claimed for your use case.
- Download the evaluation set. It includes prompts, the judge prompt and grading scripts:
hf download NousResearch/Propaganda_Eval --repo-type dataset --local-dir ./propaganda_eval - Serve both models (Arete and base Qwen3.8-27B) with the same quant, temperature and thinking setting.
- Run the prompts through both and grade them with the included judge prompt.
- Swap in a second judge model to see whether the propaganda rates hold with a different grader.
- Add your own prompts in areas Nous did not cover, and run your usual coding tests to check for regressions.
The training pairs are also public as NousResearch/arete-data-pairs, so you can read exactly what the model was taught.
Troubleshooting Arete-27B
“unknown model architecture” or the model fails to load
Your llama.cpp, Ollama or LM Studio build is too old for Qwen3.8’s hybrid Gated DeltaNet layers. Update to llama.cpp b12174 or newer, or the latest Ollama and LM Studio.
CUDA out of memory on a 16 GB card
Q4_K_M (17.2 GB) is bigger than 16 GB. Use IQ4_XS or Q3_K_L, lower the context with -c 8192, or offload a few layers to RAM. Our CUDA out of memory fix guide covers every option.
Replies start with a long think block
That is the default thinking mode. Turn it off with enable_thinking: false as shown above, or set reasoning_effort to low.
Image input does nothing
The vision projector is missing. Use -hf in llama.cpp, or pass --mmproj with one of the mmproj files from the GGUF repo.
Answers still deflect on some topics
Nous says it did not fix every behaviour. The biggest gains were in freedom of expression; other categories moved less. Try thinking on, which scored slightly higher in Nous’s provisional runs.
My Take
The interesting part is not the politics, it is the cost. Nous claims a 27B model’s political framing can be reworked with about 53 million tokens of plain fine-tuning, which is small for a lab and within reach of well-funded hobbyists.
That cuts both ways. If values are this “shallow”, anyone can push an open model in any direction just as cheaply. Publishing the eval and data is the right move, because it lets others test the claim instead of trusting it.
For most home users on 16-24 GB GPUs, Arete is a free swap for Qwen3.8-27B with the same skills. I would still wait for an independent re-run of the propaganda test before treating 0.9% as settled.
Frequently Asked Questions
What is Arete-27B?
It is Nous Research’s fine-tune of Alibaba’s Qwen3.8-27B, trained to give fewer state-aligned answers on political and value topics while keeping the same general skills.
Is Arete-27B free to use commercially?
Yes. The model card lists Apache 2.0, the same licence as Qwen3.8-27B.
How much VRAM do I need for Arete-27B?
About 16 GB for the IQ4_XS GGUF (15.2 GB) with a small context, or 24 GB for Q4_K_M (17.2 GB) with room to spare.
Is Arete-27B an uncensored model?
No. Nous only targeted political and value topics. It is not described as abliterated, and its handling of other refused requests is not documented.
Is Arete-27B better than Qwen3.8-27B at coding?
Not meaningfully. Nous reports 70% vs 67% on Terminal-Bench-2.1 but says it does not take credit for that gain.
Can Arete-27B see images?
Yes. It keeps Qwen3.8’s vision encoder, which was frozen during training. Load the mmproj file in llama.cpp to use it.
Is Arete-27B in the Ollama library?
Not as an official entry at launch. Run it with ollama run hf.co/bartowski/NousResearch_Arete-27B-GGUF:Q4_K_M.
Who judged the propaganda scores?
An LLM judge, Grok 4.7, using a rubric Nous checked against 25 hand-graded prompts. No independent test exists yet.
Sources
- Hugging Face model card: NousResearch/Arete-27B (training, results, licence)
- Nous Research blog: “Arete” (nousresearch.com/arete, 6 November 2026)
- Hugging Face: bartowski/NousResearch_Arete-27B-GGUF (quant sizes, commands, mmproj)
- Hugging Face: Qwen/Qwen3.8-27B model card (architecture, context, sampling settings, thinking controls)
- Hugging Face datasets: NousResearch/Propaganda_Eval and NousResearch/arete-data-pairs
- MarkTechPost: “Nous Research Releases Arete-27B” (7 November 2026)
- explainx.ai: Arete-27B claimed vs verified analysis (November 2026)
