Arete-27B

NousResearch's Arete-27B is a 27-billion parameter dense model post-trained directly on Alibaba's Qwen3.8-27B. Released on November 6, 2026 under the permissive Apache 2.0 license, the model specifically addresses ideological framing and political deflections embedded within base Chinese open-weight releases. By applying targeted supervised fine-tuning (SFT) using distillation from Inkling, Nous stripped state-aligned narrative framing without degrading the baseline model's core intelligence, tool use, or reasoning capabilities.

For developers seeking a capable local AI model with 27B parameters, Arete-27B occupies a crucial sweet spot. It matches the upper tier of open-weight reasoning performance while offering an uncompromised, Western-aligned baseline for autonomous agents, research, and sensitive text analysis. Because the post-training modified language weights without altering the underlying architecture, Arete-27B functions as a drop-in replacement for Qwen3.8-27B in any existing local deployment stack.

The model ships with full text-and-vision capabilities and supports an expansive 262,144-token context window. At a native BF16 weight size of 53.81 GB, it scales down comfortably to fit single consumer GPUs and mid-tier Apple Silicon setups using standard 4-bit and 5-bit quantizations.

Arete-27B maintains the exact architectural footprint of its base model, Qwen/Qwen3.8-27B (utilizing the qwen3_5 architecture class). It features a dense structure totaling 27,356,728,560 parameters distributed across 64 transformer layers.

• Attention Mechanism: Implements Grouped-Query Attention (GQA) with 4 key-value heads and a head dimension of 256. This drastically reduces the KV cache memory footprint across extended sequence lengths compared to standard multi-head attention.
• Context Window: Supports up to 262,144 tokens (256k). A full_attention_interval setting of 4 balances memory usage and long-range retrieval fidelity.
• Multimodal Stack: Retains the base model's vision-language encoder. During NousResearch's fine-tuning runs, all vision weights were frozen while language model weights were updated using text-only synthetic pairs. The Multi-Token Prediction (MTP) head present in the base architecture was disabled during training.
• Post-Training Recipe: Supervised fine-tuning across 1,202 steps (2 epochs, batch size 48) on roughly 52.8 million tokens with a 64k native training context. Nous utilized AdamW with bf16 compute and fp32 master weights, omitting sequence packing, reinforcement learning, and direct preference optimization.

Because dense 27B parameters require continuous activation during inference, the compute pattern is predictable and consistent. You avoid the dynamic memory routing overhead of Mixture-of-Experts architectures, translating to uniform token generation latencies across long prompts.

Arete-27B excels at instruction-following, complex chat interactions, structured data generation, and systemic reasoning. The primary justification for deploying Arete-27B over other models in its weight class centers on factual integrity and viewpoint independence.

Base Chinese models often refuse queries regarding domestic civil liberties, state surveillance, or territorial disputes, or they default strictly to party-sanctioned framing. Arete-27B was explicitly tuned across seven critical domains: freedom of expression, state ideology, market economics, education, surveillance and data privacy, democracy, and environmental policy.

On Nous's 803-prompt Propaganda_Eval benchmark, propaganda-aligned responses dropped from 22% in the base model to 0.9% in Arete-27B, while the factual accuracy score increased from 2.32 to 2.98 out of 5. For automated document review, policy research, and historical data extraction, it delivers unvarnished factual recall without deflection.

With 262k native context support and refined instruction-following, Arete-27B is engineered for multi-step agent frameworks such as Nous Hermes Agent. It handles deep repository analysis, automated code synthesis, and multi-document synthesis without losing track of system prompts or tool-calling schemas.

Because the underlying vision encoder remains intact, the model handles mixed visual-text inputs. You can ingest high-resolution diagrams, complex financial charts, technical schematics, and mixed PDFs locally while relying on the language layers for uninhibited textual interpretation.

To run Arete-27B locally, hardware planning hinges primarily on managing weights and KV cache memory. At 27B parameters, memory consumption is too large for 16GB VRAM GPUs unless heavily quantized, but it runs efficiently on 24GB cards and unified memory platforms.

The unquantized BF16 weights take up 53.81 GB, necessitating multi-GPU setups for full-precision local inference. Quantization via llama.cpp (GGUF) or MLX makes single-device execution practical:

• Best GPU for Arete-27B: An Nvidia GeForce RTX 4090 or RTX 3090 (24GB VRAM) represents the best consumer GPU choice. Running the Q4_K_M quantization leaves roughly 5GB to 6GB of VRAM headroom for the context cache, comfortably handling up to 16k to 32k tokens on a single card without offloading to system RAM.
• Apple Silicon Execution: For macOS users, MLX 4-bit runs with exceptional efficiency. An M4 Pro or M3/M4 Max with at least 36GB of unified memory allows the model to process up to 64k tokens locally with zero swap overhead.
• Dual GPU Builds: Users with dual RTX 3060 12GB cards or dual RTX 4060 Ti 16GB cards can split Q5_K_M or Q8_0 weights cleanly across both buses via PCIe peer-to-peer or vLLM tensor parallelism.

Arete-27B performance varies based on hardware bandwidth:

• RTX 4090 (Q4_K_M via llama.cpp / vLLM): Expect roughly 38 to 48 tokens per second for raw generation, with rapid prompt processing exceeding 1,800 tokens per second.
• Apple M4 Max (MLX 4-bit): Achieves approximately 32 to 38 tokens per second.
• Dual RTX 3090 (Tensor Parallelism TP=2, Q8_0): Delivers 40 to 50 tokens per second with ample headroom for full 64k contexts.

The simplest path to launch the model locally is using Ollama or llama.cpp:

1# Pull and run the community GGUF release
2ollama run hf.co/bartowski/Arete-27B-GGUF:Q4_K_M

For extended context sessions, configure the num_ctx setting in your Modelfile to allocate the necessary KV cache:

1FROM hf.co/bartowski/Arete-27B-GGUF:Q4_K_M
2PARAMETER num_ctx 32768
3PARAMETER temperature 0.7

The comparison here is direct, as Arete-27B derives entirely from Qwen3.8-27B. Compute requirements, layer dimensions, context limitations, and raw inference speed are identical. The difference lies strictly in output distribution.

Qwen3.8-27B frequently exhibits defensive deflection, refuse patterns, or pro-CCP narratives on questions touching Chinese governance, cross-strait relations, or state regulation. Arete-27B removes this political filtering. In general reasoning benchmarks and synthetic coding evaluations, Arete-27B performs at parity with the base model, proving that alignment fine-tuning did not introduce catastrophic forgetting.

Google's Gemma 2 27B is the standard benchmark for non-MoE models in the high-20B tier. Gemma 2 features a 42-layer architecture with an 8k native sliding-window context, which severely restricts long-document analysis and multi-turn agent history compared to Arete-27B's 262k context.

While Gemma 2 27B offers strong raw language modeling, its strict context limits and aggressive pre-training safety filters make it less versatile for complex, long-context system tasks. Furthermore, Gemma 2 runs under the Gemma Terms of Use, whereas Arete-27B provides full commercial freedom under Apache 2.0.

Qwen 2.5 32B requires slightly more VRAM (roughly 20GB for Q4_K_M compared to Arete's 17.2GB), which pushes 24GB GPUs close to memory capacity when using extended contexts.

While Qwen 2.5 32B remains a standout code and math model, Arete-27B operates with lower inference overhead, fits more reliably on single 24GB cards with substantial context buffers, and eliminates the ideological constraints inherent in Alibaba's native checkpoints.