Nous Research releases Arete-27B — a 27B open fine-tune of Qwen3.8-27B that runs on one GPU

What was released, from Nous Research's announcement and the Arete-27B model card on Hugging Face, both read on Nov 7, 2026: Nous fine-tuned all of Qwen3.8-27B's language-model weights (the vision encoder was left frozen; there was no reinforcement learning, preference optimisation or architecture change) on roughly 52.8 million tokens over two epochs and 1,202 steps, using Thinking Machines' Inkling as the teacher. The weights, the Propaganda_Eval dataset and the arete-data-pairs training data are all on Hugging Face under the Apache 2.0 licence, per the model card.

The results are Nous's own, and are labelled as such here: on its 803-prompt evaluation across seven value categories, judged by an LLM (Grok 4.7) against Nous's own rubric and not independently reproduced, answers flagged as containing propaganda fell from 22% for the base model to 0.9% (thinking off, temperature 0), and its fact score rose from 2.32 to 2.98 out of 5. Nous reports no meaningful regression on its Hermes Bench (65.77% to 65.60%) and a Terminal-Bench 2.1 move from 67% to 70% that it explicitly does not take credit for, citing harness differences. Treat the headline numbers as the maker's measurements until someone outside Nous reproduces them.

The renter maths is the practical story. At 27 billion parameters the weights are about 54GB in BF16 and about 27GB at 8-bit, before KV cache and headroom — one 48–80GB card, not a cluster. On today's checked board (Nov 7, 2026) that means an A100 80GB at Thunder Compute, $1.09 an hour, holds it in BF16, and an L40S at RunPod Secure, $1.09 an hour, holds the 8-bit form with room to spare; a 24GB card such as an RTX 4090 takes it quantised. A behaviour change that frontier labs describe in pretraining terms was here done with about 53 million tokens of fine-tuning — an afternoon-scale job on one rented card.