Nous Research Releases Arete-27B — a Post-Trained Version of Alibaba's Qwen3.8-27B

Nous Research published the Arete-27B model on HuggingFace for the text-generation task. It is a post-trained version of Alibaba Qwen3.8-27B: the authors performed supervised fine-tuning to adjust the model's responses on politically sensitive topics. According to Nous Research's estimate, the share of responses containing propaganda dropped from 22% to 0.9%, while the overall factual score rose from 2.32 to 2.98 out of 5.

What it means

The model contains 27.4 billion parameters; the weights are posted in safetensors format and take up 51 GB. The license is apache-2.0, meaning the model can be used and modified freely, including in commercial projects. The largest weights file is 3.7 GB. For running on your own hardware, quantized builds are available: there are eight of them on the site in GGUF and EXL formats. The smallest working build weighs 13.1 GB — IsValorum/Arete-27B-VAL-APEX-I-MiniPlus-V3-GGUF. That is a level that fits into the VRAM of modern consumer GPUs, although the full model in its original form will require noticeably more resources. The developer does not specify the VRAM requirements. The model is trained only on text data with a 64k context, without reinforcement learning. The authors emphasize that they have not eliminated all the quirks of the original Qwen3.8-27B, but believe they have addressed one of the possible causes of the observed biases. You can follow news about tools for local deployment in the CPU3D news feed.