A glowing desktop setup displays an AI interface for chat, document retrieval, and source verification.
Local AI usually gets judged by speed and size. Tokens per second, parameter counts and whether a quant fits in VRAM dominate the conversation. XDA Developers' Abhinav Raj took a different approach. He put a retrained Qwen3.8 variant through an everyday privacy question on a 16GB gaming GPU, then compared it with the stock model.

The retrained model is Arete-27B, from Nous Research. The headline says the censorship was "trained out" and that the model "stopped overthinking." Both claims go further than the evidence supports. It's still a useful look at what a targeted fine-tune does on a PC.

A glowing desktop setup displays an AI interface for chat, document retrieval, and source verification. What Arete-27B actually is​

Nous says it post-trained Alibaba's Qwen3.8-27B to reinforce western values on a Chinese-trained model, and calls the result Arete-27B. The model card lists the method as supervised fine-tuning (SFT), with Inkling as the teacher model and text-only post-training data. The card also says training used about 52.8 million tokens at 64k context, over two epochs and 1,202 steps.

Nous published the weights, the evaluation set and the training data. XDA reports an Apache 2.0 license. Some third-party write-ups say the license is unpublished, so check the Hugging Face model card before you ship it in anything.

This is not the usual "abliterated" build. XDA notes that abliterated models strip refusals across the board. Arete is a targeted retrain across seven categories, from free expression to data privacy and surveillance. Calling it "censorship trained out" is shorthand. Nous describes a change in behavior on specific topics, not a blanket removal of refusals.

Nous's own numbers​

Nous measured results on its own 803-prompt evaluation, scored by an LLM judge. Treat them as vendor figures.

Measure (thinking off)Qwen3.8-27BArete-27B
Fact score (0–5)2.322.98
Answers with propaganda22%0.9%
Hermes Bench (internal)65.77%65.60%
Terminal-Bench 2.167%70%

The Hugging Face model card lists the same figures for fact score, propaganda rate, Hermes Bench and Terminal-Bench 2.1. At temperature 0.6 the card reports 2.88 for Arete and 2.24 for Qwen.

Three caveats apply:

  • The judge was chosen by Nous. One outside write-up, from explainx.ai, points out that the 22%-to-0.9% figure is Nous's claim as scored by a judge Nous selected.
  • The benchmark only covers politically sensitive prompts. It says nothing about everyday accuracy or phone-privacy advice.
  • Hermes Bench is internal. Nous says the 0.17-point gap is within noise. That's reassuring, but it isn't an independent audit.

The hardware setup​

Raj ran both models in llama-server on an RTX 4070 Ti Super, a Ryzen 5 7600X and 32GB of DDR5-6000. The default Q4_K_M files don't fit in 16GB of VRAM. He used bartowski's IQ4_XS quants instead, at 15.2GB and 15.5GB. That left about 2GB for everything else, including context.

Both models used the thinking-mode samplers from Qwen's model card: temperature 1.0, top_p 0.95, top_k 20 and min_p 0.

This setup is tight. With so little headroom, context length becomes the limiting factor, and that shaped the search test.

Test one: no web search​

Raj asked both models which they'd pick, a Xiaomi phone or a Google Pixel, for a privacy-minded buyer. He also asked what data each company collects, where it ends up and who could access it.

  • Verdict: Both picked the Pixel.
  • Framing: Arete called Chinese state access a structural risk with no real Google equivalent. Qwen spent more time on US surveillance. It warned that a Pixel wouldn't help much if that was the buyer's main worry.
  • Facts: Neither model got it right. Qwen listed more of the regions where Xiaomi says it stores data, while Arete stopped at "China and/or other countries." Qwen mentioned a "permissionless location" setting that Raj says Google's Pixel pages don't mention. Arete mixed up two different Pixel privacy features.
  • Reasoning length: Arete reasoned for 3,371 words. Qwen used about 6,000.

Qwen's answer arguably gave the fuller picture of Western surveillance risk, even though its facts slipped. A bias correction doesn't make a model more accurate. It changes which risks the model emphasizes.

Test two: with web search​

Raj then connected both models to a self-hosted SearXNG instance. He sent the same prompt, plus a request to search and cite sources, starting at 32K context.

Arete-27BQwen3.8-27B
Searches912
Pages read412
Page content cappedYes, every pageNo
ContextAbout 22K tokens (fit in 32K)Ran out at 32K; finished at 64K with 8-bit cache
TimeAbout 15 minutes38 minutes

Qwen lost time chasing a Romanian fine against Xiaomi and a 404 Media story it half-remembered and never found. Even so, Raj judged its answer the more careful one. It read Xiaomi's privacy policy, reported which sites blocked it and dropped claims it couldn't confirm.

Arete was quicker but sloppier on sourcing:

  • It applied a study of phones sold in China to every Xiaomi device.
  • It cited a figure from a page it never really read.
  • It leaned on a forum post.

Qwen made errors too. Raj says it overstated what US national security letters can obtain from Google. Both models sounded more certain after searching than before, which is a bad trade if the sources are shaky.

What this does and doesn't show​

Raj's own conclusion is measured. Across one run of each test, Arete thought less, searched less and fit in 32K of context. It also checked its sources less carefully. He'd use it for everyday questions and double-check anything it finds online.

Four limits matter here:

  1. Single runs. Sampling at temperature 1.0 varies between runs. The word counts and timings may not repeat.
  2. Confounded variables. Quantization, page-capping behavior and the search setup all affect the outcome. The test doesn't isolate which factor caused which difference.
  3. Not a Xiaomi or Google audit. The privacy claims are the models' outputs, as fact-checked by one author.
  4. Not proof that Arete "stopped overthinking." Shorter reasoning here may reflect training on a teacher's more concise answers. It could also mean the model is less thorough. This test can't tell those apart.

Practical takeaways for local AI users​

  • Mind your VRAM. On a 16GB card, a 27B model at IQ4_XS leaves little room. If an agentic search workflow overruns 32K, you either raise the context and quantize the cache or lose the run. Qwen needed exactly that here.
  • Cap what tools pull in. Arete's per-page character limit kept it inside the context window. If your front end lets you cap fetched content, it may help.
  • Treat training lineage as a feature. If you use a local model for research on sensitive topics, know which way its priors lean. A fine-tune can shift that lean without making the model fact-check any better.
  • Verify privacy claims at the source. For anything like a phone-privacy decision, read the vendors' privacy policies. Don't rely on either model's summary.

Raj's test is a small sample, but it makes the right point. A local model's training matters as much as its tokens per second, and a fine-tune that fixes one bias can leave other weaknesses in place.

 

References

  1. I ran Qwen3.8 with its censorship trained out, and it stopped overthinking XDA 2026-11-20T11:00:21+00:00
  2. NousResearch/Arete-27B · Hugging Face huggingface.co