27B dense (Qwen3.8-27B fine-tune)
262,144 tokens native
Apache 2.0
Nous Research post-trained Alibaba's Qwen3.8-27B to alter refusals, deflections, and Chinese-state framing that it observed on politically sensitive topics, describing the objective as reinforcing Western values while demonstrating that open-weight models can be adapted to different value systems. Arete retains the dense Qwen3.8-27B hybrid architecture and native 262,144-token context window, with 64 language layers arranged as repeated groups of three Gated DeltaNet layers followed by one full-attention layer. It retains the base model's image and video encoder, although the Arete post-training data was exclusively text and the vision encoder remained frozen. Nous used supervised fine-tuning only, with no reinforcement learning, preference optimization, or intended architectural modification. The training corpus was generated from 20,000 deduplicated prompts spanning freedom of expression, state ideology, state versus free markets, education, data privacy and surveillance, democracy, and environmental policy. Inkling served as the open-weight teacher; overlapping prompts were removed against the held-out evaluation set, and Nous publicly released the training data and pipeline artifacts. Fine-tuning covered approximately 52.8 million tokens at 64K training context without sequence packing, using two shuffled epochs, 1,202 steps, batch size 48, AdamW, BF16 computation with FP32 master weights and optimizer states, and tensor parallelism of four. All language-model weights were trained while the vision encoder was frozen. The Multi-Token Prediction head was disabled during training, and the released configuration contains no MTP layers. Nous's 803-prompt Propaganda_Eval uses English-language statutes, news, and research sources, with Grok 4.7 acting as a blinded judge over required facts and official-frame language. At temperature zero with thinking disabled, Nous reports that Arete increased the mean factual score from Qwen3.8-27B's 2.32 to 2.98 out of five and reduced answers containing measured propaganda from 22% to 0.9%; answers receiving zero factual points fell from 359 to 216. The largest category change was freedom of expression, where the factual score increased from 0.89 to 2.44 and the measured propaganda rate fell from 60% to 1%. General-capability checks showed Hermes Bench moving from 65.77% to 65.60%, which Nous interprets as noise, while Terminal-Bench-2.1 moved from 67% to 70%; Nous explicitly does not claim credit for the latter increase. The checkpoint preserves Qwen's modified ChatML-style template, optional thinking mode, system prompts, multimodal message rendering, and structured tool calling. Default generation metadata specifies temperature 1.0, top-p 0.95, and top-k 20. The Apache-2.0 checkpoint contains 18 safetensors shards totaling approximately 54.7 GB.