Arete-27B: Nous Research's Retuned Qwen3.8-27B and Its Requi

Arete-27B is an open-weight language model released by Nous Research on November 6, 2026 (UTC). It is Alibaba's Qwen3.8-27B with supervised fine-tuning (SFT) only, aimed at reducing the refusals, deflections and Chinese-state-framed answers the base model gave on politically sensitive topics. It is licensed under Apache-2.0. Because it shares the base model's architecture, memory needs are the same: about 54.7GB for the BF16 weights and about 17.2GB for a Q4_K_M GGUF (actual community file size), so a 4-bit build fits a 24GB GPU or a 32GB Mac.

In Nous Research's own evaluation (803 prompts, LLM judge), the fact score moved from 2.32 to 2.98 out of 5, and the share of answers containing propaganda fell from 22% to 0.9%. Terminal-Bench-2.1 went from 67% to 70%, and its in-house Hermes Bench stayed about level (65.77% to 65.60%). All of these are developer-reported numbers, and the direction of the value adjustment is Nous's own definition. This article keeps that distinction while covering specs, memory requirements, how to run it, and cautions for business use.

- What it is: Qwen3.8-27B (dense, ~27B) re-tuned with SFT only. No RL, no preference optimization, no architecture change. Apache-2.0
- Difference from the base: improved scores on a custom eval of factual recall and adoption of the official Chinese-state line across seven political categories; general and agentic benchmarks self-reported as level
- Memory: BF16 ~54.7GB / Q8_0 ~28.7GB / Q4_K_M ~17.2GB (weights only; KV cache is extra)

According to Nous Research, Qwen3.8-27B produced refusals, deflections and answers framed in line with the Chinese government on politically sensitive topics. Arete-27B is the result of further training to move responses toward what the company calls Western values. The name comes from the Greek arete (excellence, virtue). The official blog title says the model was post-trained to reinforce Western values, so the value direction is stated explicitly.

It covers seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. The model card says it is intended for general assistance and agentic work by users who want this shift in viewpoint, and for research into how post-training changes a model's values and behavior.

The model card lists the training settings in detail (AdamW, learning rate 1e-5 cosine-decayed to 1e-6, batch size 48, tensor parallel 4, and so on). Training loss fell quickly during the 120-step warmup, leveled off near 1.0 by the end of epoch 1, and stepped down to about 0.8 at the start of epoch 2. The card explains the step as the model seeing the same examples a second time.

The core is Nous's own Propaganda_Eval, 803 prompts built from English-language government statutes, news and research reports. Each prompt has 2 to 4 required facts taken from a dated source article, plus the official-frame language the CCP uses on that topic. An LLM judge, Grok 4.7, scores factual recall (0 to 5) and framing (0 to 2; "any propaganda" is the share of answers scoring 1 or 2). The judge is not told which model wrote the answer.

These are Nous Research's published figures (vendor-measured). Each row is the mean of five runs over all 803 prompts, and ± is the run-to-run standard deviation. A caveat on the card: Arete-27B and base Qwen3.8-27B ran at temperature 0, while the other rows used 0.6, so decoding settings are not identical across models. At temperature 0.6, Arete-27B scores 2.88 ± 0.03 and Qwen3.8-27B 2.24 ± 0.05. The blog also lists Arete-27B with thinking on at 3.07 / 0.6%, but marks it provisional (2 of 5 runs).

Answers scoring 0 fell from 359 to 216 of 803, and the improvement held in all seven categories. The largest change was freedom of expression: the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%. Note also that the absolute fact score is still below the teacher Inkling (3.96), Nemotron 3 Ultra and Mistral Large 4. Reducing adoption of the official line and breadth of factual coverage are different measures.

The Hermes Bench gap is 0.2 points, which Nous itself reads as noise. For the 3-point Terminal-Bench-2.1 difference, the company says it does not take credit, noting that both models were run in a like-for-like configuration on its own hardware and that results can differ from official benchmarks. The safe reading is "no capability regression," not "improved."

Arete-27B's config.json matches Qwen3.8-27B in architecture (Qwen3_5ForConditionalGeneration), 64 layers, hidden size 5,120, 4 KV heads, head_dim 256, and the hybrid layout with full attention every fourth layer (the rest are linear attention). The differences are that the MTP head layer count is 0 (the base has 1) and the total weight count is about 27.36B (the base is about 27.78B). The maximum context (262,144 tokens) is the same. Memory can therefore be estimated almost exactly as for the base; see Qwen3.8-27B requirements.

If you want to check quickly whether it runs on your GPU or Mac, the VRAM calculator (free, no sign-up) estimates the required VRAM from the model, quantization and context length. Arete-27B has the same structure as the base Qwen3.8-27B, so picking Qwen3.8-27B in the list gives the same figures.

GGUF sizes are taken from the actual files in bartowski's quantizations (bartowski/NousResearch_Arete-27B-GGUF), with GB meaning 10^9 bytes. These are third-party quantizations, not an official Nous Research release. The table shows weights only; KV cache and runtime buffers come on top.

Of Arete-27B's 64 layers, only 16 use full attention; the rest use linear attention with fixed-size state. The KV cache grows only for those 16 layers. From the config, that is 16 layers x 4 KV heads x head_dim 256 x 2 (K and V) x 2 bytes (FP16/BF16) = about 64KB per token (estimate).

- 64k tokens (the training context): ~4GB
- 128k tokens: ~8GB
- 262,144 tokens (maximum): ~16GB

This excludes the state held by the linear-attention layers and runtime buffers, so budget a few extra GB. Quantizing the KV cache to 8-bit roughly halves it. Compared with a standard-attention dense model such as Nanbeige4.3-33B, this hybrid layout keeps KV growth small even at long context.

The model card does not include inference commands or recommended sampling settings. The shipped generation_config.json uses temperature 1.0, top_p 0.95 and top_k 20, which matches what the base Qwen3.8-27B recommends for thinking mode (temperature 1.0, top_p 0.95, top_k 20, presence_penalty 0.0). The base has thinking on by default, tunable with reasoning_effort, and recommends temperature 0.7, top_p 0.80 and presence_penalty 1.5 when thinking is off. Applying the same settings to Arete-27B is natural, but Nous itself has not stated these as recommendations.

For llama.cpp, the quickest route is the one in the quantized repo's README. The bartowski README says llama.cpp b12174 or newer is required, and fetching with -hf also pulls the mmproj for image input automatically.

The same GGUFs also load in GGUF-capable apps such as LM Studio, koboldcpp and Jan (the bartowski README says they work there). Several community members have published 4-bit, 5-bit and 8-bit MLX builds, which are another option on Apple Silicon. As for Ollama, as of November 9, 2026 we could not confirm any official listing from Nous Research or the Ollama library; you would import a GGUF yourself.

To run BF16 on vLLM or SGLang, the recipes for the base Qwen3.8-27B are a useful reference. Because the architecture is identical, swapping the model name is the basic approach, but no Arete-specific procedure is published. FP8, W4A16 and NVFP4 builds exist from community members; we have not verified their quality.

The difference from the base Qwen3.8-27B is the direction of behavior, not weight size. Memory needs are nearly identical, so hardware fit does not separate the two. The other two models are separate lines, with figures in our own articles (the Nous Hermes guide and Nanbeige4.3-33B requirements). Arete-27B is not a successor to Hermes; it is a derivative of Qwen3.8-27B.

The teacher, Inkling, is a large open-weight MoE from Thinking Machines; see Inkling requirements and the Inkling 2 Small overview. The 27B model is positioned as a distillation of its answers, though the fact score is 2.98 against the teacher's 3.96, so a gap remains.

- License: Arete-27B is Apache-2.0 and the base Qwen3.8-27B is Apache-2.0, so there is no extra scale restriction on commercial use or redistribution. Community quantizations and "uncensored" derivatives need their own license and provenance checks
- What "value adjustment" means: the adjustment is a shift toward what Nous Research calls Western values, not a guarantee of neutrality or correctness. The evaluation criteria (choice of required facts, definition of the official line, judge design) were also set by the company; third parties can reproduce them with the published eval set
- Vendor-measured numbers: fact score, propaganda rate, Terminal-Bench-2.1 and Hermes Bench are all developer-reported, and the judge is an LLM (Grok 4.7). The card says the rubric was revised three times against a 25-prompt pilot graded by hand
- Japanese performance is not evaluated: the eval prompts come from English-language sources, and neither the card nor the blog covers factuality or bias in Japanese. For Japanese use, build a small evaluation set and check yourself
- Bias is not fully removed: Nous itself says it does not claim to have remedied all related behaviors. The remaining 0.9%, and topics outside the evaluation, may still show base-model tendencies
- Little difference for general tasks: for work that avoids political topics (code generation, internal document summaries), performance is within the self-reported range of the base. If you have no reason to prefer it, there is no need to switch
- Image input: the vision encoder was frozen but remains, and GGUF builds ship an mmproj. Behavior on image input is not officially evaluated

- Hugging Face model card: https://huggingface.co/NousResearch/Arete-27B
- Nous Research blog post: https://nousresearch.com/arete
- Evaluation dataset: https://huggingface.co/datasets/NousResearch/Propaganda_Eval
- Training data: https://huggingface.co/datasets/NousResearch/arete-data-pairs
- Base model: https://huggingface.co/Qwen/Qwen3.8-27B
- Actual GGUF sizes (bartowski): https://huggingface.co/bartowski/NousResearch_Arete-27B-GGUF