Nous Research retrains Qwen3.8 into Arete-27B
The author tested Arete-27B, a retrained version of the Qwen3.8-27B model developed by Nous Research to remove politically sensitive biases and refusal patterns. Running the models locally on a 16GB GPU revealed that Arete-27B reasoned with significantly fewer words, finished tasks faster, and used less context than the original model, though the author notes it requires careful source verification when using web search. Open article →