Nous Research has released Arete-27B, a 27-billion parameter model derived from Alibaba’s Qwen3.8-27B. The fine-tune reduces the frequency of Chinese state-aligned answers on sensitive topics from 22% to 0.9%. The weights, training data, and evaluation scripts are available publicly under Apache 2.0 and MIT licenses.
Specs and performance
The model uses 27 billion dense parameters and trains at a 64k context window. The base Qwen model supports up to 262,144 tokens natively. BF16 weights occupy 53.81 GB, while a community quantised version fits in 17.20 GB, targeting 24 GB consumer GPUs.
Capability scores remain stable on the benchmarks Nous tested. Political framing now aligns with Western models. The propaganda rate dropped to 0.9%, down from 22% for the base model. Fact recall improved to a score of 2.98 out of 5, though 216 of the 803 test answers still received a score of zero.
Training method
Arete-27B is a supervised fine-tune only. Nous applied no reinforcement learning, preference optimisation, or architectural changes. The team chose Qwen3.8-27B as the base because it was the most capable local model and displayed the specific behaviour they wanted to correct.
The training teacher was Inkling, an open model from Thinking Machines. Nous selected an open teacher to avoid terms-of-use restrictions on distillation. The team rejected Nemotron 3 Ultra due to inconsistent performance on China-related topics.
Training processed approximately 52.8 million tokens at a 64k context window without sequence packing. The run took two epochs and 1,202 steps using a batch size of 48. It used AdamW with a peak learning rate of 1e-5 and tensor parallelism of 4.
All language model weights were trained. The vision encoder remained frozen because the data is text-only. The multi-token prediction head was disabled. The released arete-data-pairs dataset contains 34,878 prompts across seven topics. Each prompt includes both an Inkling answer and a Qwen answer. Twenty-nine thousand two hundred and thirty-six accepted teacher answers form the supervised fine-tuning split.
Evaluation details
The Propaganda_Eval set comprises 803 prompts across seven value categories, including freedom of expression, state ideology, data privacy, and democracy. Each prompt requires two to four specific facts from a cited source.
A Grok 4.7 judge grades two metrics separately:
- Fact mean (0 to 5): measures how many required facts the answer states. An answer opposed to a fact scores zero.
- Any propaganda: calculates the share of answers that adopt the state’s account as their own.
The judge must quote the exact phrase to flag propaganda. Nous removed every training prompt that overlapped the evaluation set. The eval is MIT licensed and includes grading scripts.
Benchmark results
On the model card page, Arete-27B raised the fact mean from 2.32 to 2.98. Propaganda fell from 22% to 0.9%. Answers scoring zero dropped from 359 to 216 of 803.
The largest change occurred in freedom of expression. Fact score rose from 0.89 to 2.44, and propaganda fell from 60% to 1%.
Capability held on Nous’s tests. Hermes Bench moved from 65.77% to 65.60%, which Nous describes as noise. Terminal-Bench-2.1 rose from 67% to 70% in Nous’s harness. Nous does not claim credit for that gain.
Comparison table
| Metrics | Arete-27B | Qwen3.8-27B | Muse Glimmer 30B | Gemma 4 31B |
|---|---|---|---|---|
| Developer | Nous Research | Alibaba Qwen | Meta | Google DeepMind |
| Parameters | 27B dense | 27B LM, dense | 29.6B incl. vision, dense | 30.7B, dense |
| Context | Not disclosed (trained at 64k) | 262,144 native, up to 1M | 131,072+ | 256K |
| License | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
| Input | Text, image (encoder frozen) | Text, image, video | Text, image | Text, image |
| 4-bit footprint | 17.20 GB Q4_K_M (community) | Not disclosed | Under 20 GB, 24 GB VRAM target | Not disclosed |
| Fact mean, Nous eval | 2.98 | 2.32 | 3.02 | 2.58 |
| Any propaganda, Nous eval | 0.9% | 22% | 0.5% | 0.5% |
| Terminal-Bench 2.1 | 70% (Nous harness) | 67% Nous harness; 73.0 official | 51.7 (Meta, terminus2) | Not disclosed by Google |
| GPQA Diamond | Not disclosed | 89.2 | 83.5 | 84.3 |
Note: Arete and Qwen ran at temperature 0 in the Nous eval. Muse Glimmer and Gemma ran at 0.6. Terminal-Bench figures come from different harnesses.
Deployment
Weights ship in BF16 on Hugging Face. No hosted inference provider serves the model yet. Community GGUF quants run in llama.cpp, LM Studio and Ollama, ranging from 8.64 GB to 53.81 GB. The GGUF build also includes the image projector.
What it means
Developers can now run a Qwen-class model locally without accepting state-aligned answers on sensitive topics. The fix required only 52.8 million tokens of plain supervised fine-tuning. Hermes Bench held at 65.60% versus 65.77% for the base model. Fact recall rose to 2.98 of 5 but still trails peers. Weights, eval and training data are open under Apache 2.0 and MIT.
