This is a plan, not a write-up. I haven’t run it yet. After reading about Nous Research’s Arete-27B, I want to compare it with the base model on my own prompts.
The plan
- Pull a Q4_K_M GGUF of Arete-27B (about 17.2GB) from Hugging Face into Ollama, and pull stock Qwen3.8-27B at the same quant so the comparison is fair.
- Check that both fit in VRAM with my usual context length, and write down tokens per second for each.
- Run the same prompt set through both: everyday coding and shell questions, a few agentic tool-use tasks, and some history and current events questions where framing matters.
- Score answers side by side for accuracy, refusals and tone, without reading too much into a small sample.
- Record the model versions and quant hashes so the results can be reproduced.
Questions I want answered
- Does Arete keep Qwen’s coding and agent skills at Q4?
- How different do the answers look on topics where framing matters?
- Is the speed identical, as it should be for the same architecture?
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
Isolate it: pull models into a separate test setup first, and keep a note of what you downloaded so you can remove it cleanly.
God Bless, and May the Source Be with You!
