Arete-27B Text Generation Model: 12GB+ VRAM, GGUF Builds

- 1. At a Glance
- 2. Overview
- 3. Specifications
- 4. Performance
- 5. Strengths and Use Cases
- 6. Hardware Requirements
- 7. Can You Run It Locally?
- 8. Recent Models in the Same Size Class
- 9. How to Get It
- 10. Quantized and Converted Variants
- 11. Other Models for the Same Task
- 12. What to Read Next
- 13. Sources
- 14. Update History
At a Glance
| Item | Value |
|---|---|
| Repository | NousResearch/Arete-27B |
| Family guide | Arete guide (1 articles) |
| Publisher guide | Nous Research: models and licenses |
| Published | 2026-11-07 |
| License | apache-2.0 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Nous Research has released “Arete-27B", a new text generation model that applies value alignment via supervised fine-tuning (SFT) to Alibaba’s open-weight model “Qwen3.8-27B".
It is built with the aim of allowing open-source model users to freely choose values and perspectives tailored to their own preferences. It aims to correct biases, response refusals, and propagandistic expressions on politically sensitive topics observed in the base model Qwen3.8-27B, achieving improved factual accuracy and reduced propaganda expressions. It is provided for users seeking responses aligned with specific perspectives and values while maintaining general instruction-following and agent task performance, as well as for researchers investigating behavioral changes through post-training.
Specifications
- Parameters: 27B
- Architecture: Qwen3_5ForConditionalGeneration
- Context length: 64k
Performance
In the model card published by the authors, alignment evaluation results using a proprietary evaluation dataset (803 prompts) spanning 7 value categories, alongside measurement results for general capabilities and agent performance, are made public.
The comparison table with other models in the alignment evaluation is as follows:
| Model | Fact mean (0–5) ± SD | Any propaganda |
|---|---|---|
| Inkling (teacher) | 3.96 ± 0.07 | 0.3% |
| Mistral Large 4 | 3.65 ± 0.04 | 0.3% |
| Muse Glimmer 30B | 3.02 ± 0.05 | 0.5% |
| Gemma 4 31B | 2.58 ± 0.05 | 0.5% |
| Nemotron 3 Ultra (thinking off) | 3.22 ± 0.02 | 0.6% |
| Arete-27B step1202 (thinking off) | 2.98 ± 0.03 | 0.9% |
| GLM 5.3 (thinking off) | 3.10 ± 0.06 | 15% |
| DeepSeek V4.1 Flash (thinking off) | 2.39 ± 0.05 | 16% |
| Qwen3.8-27B (starting model, thinking off) | 2.32 ± 0.03 | 22% |
Furthermore, the comparison results of general capabilities and agent performance with the base model Qwen3.8-27B are as follows:
| Benchmark | Qwen3.8-27B | Arete-27B |
|---|---|---|
| Hermes Bench (internal general-capability benchmark) | 65.77% | 65.60% |
| Terminal-Bench-2.1 | 67% | 70% |
According to the measurements published by the authors, Arete-27B improved its average fact score (Fact mean) from 2.32 to 2.98 compared to the original model, and the proportion of answers containing propaganda (Any propaganda) dropped sharply from 22% to 0.9%. Improvements were confirmed across all 7 evaluated categories, with the “Freedom of expression" area showing the largest improvement, where the fact score rose from 0.89 to 2.44 and the propaganda rate dropped from 60% to 1%. On the other hand, it does not reach the fact scores of models used as teacher models or comparisons, such as Inkling (fact score 3.96), Mistral Large 4 (3.65), and GLM 5.3 (3.10). In addition, it recorded a score of 70% on Terminal-Bench-2.1, which measures whether tasks can be completed by actually executing commands in a terminal, and the score difference on Hermes Bench (65.77% vs 65.60%) is reported to be within the margin of noise, indicating that general functions and agent performance have not degraded due to post-training.
Strengths and Use Cases
Arete-27B is designed to improve the bias, response refusals, and propaganda expressions that the original model showed in specific politically sensitive topics, and to output responses in line with the perspectives and values desired by users. Specifically, it excels at dialogue and analysis tasks related to the following 7 value categories:
- Freedom of expression
- State ideology
- State versus free market
- Education
- Data privacy and surveillance
- Democracy
- Environmental Policy
In adjusting the model, evaluation and measurement were conducted by clearly separating “fact recall"—whether verified facts can be accurately cited—and “framing"—whether a specific claim is adopted as one’s own view. The open model “Inkling" was adopted as the teacher model, and supervised fine-tuning (SFT) was performed using a dataset constructed from 20,000 non-overlapping prompts spanning the above 7 categories. In this training, a 64k context length was used, and fine-tuning was performed on approximately 52.8 million tokens of text data.
This model is intended for use in general dialogue assistant tasks that emphasize unbiased expression or responses aligned with specific values, as well as agent tasks such as handling terminal operations. Furthermore, it is useful as a base model for research purposes to investigate how post-training using only SFT affects the values and behavior of a model.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 27.4B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 12GB (RTX 4070 / 3060 12GB, etc.) | Q2_K | 9.9GB | 11.8GB |
| 16GB (RTX 5060 Ti 16GB / 4060 Ti 16GB, etc.) | Q3_K_L | 12.9GB | 15.5GB |
| 24GB (RTX 4090 / 3090, etc.) | Q5_K_M | 19.3GB | 23.1GB |
| 32GB (RTX 5090, etc.) | Q8_0 | 26.7GB | 32.0GB |
| 80GB class (A100 / H100) | BF16 | 50.1GB | 60.1GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build bartowski/NousResearch_Arete-27B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but bartowski/NousResearch_Arete-27B-GGUF provides a GGUF build you can use.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
Compression: the Q2_K build measures 3.11 bits per weight — about 19% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| LiquidAI/LFM2.5-24B-A2B-Base | 24.0B | 80GB | lfm1.0 | LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-FP8 | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-bf16 | 24.0B | 80GB | lfm1.0 | LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-8bit | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B | 24.0B | 12GB | lfm1.0 | LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ VRAM (2026-11-17) |
How to Get It
Arete-27B is distributed in safetensors format on the Hugging Face model repository (NousResearch/Arete-27B).
Since this model is not a Gated model that requires prior agreement to terms of use, it can be obtained without any procedures. The Apache 2.0 license is applied.
To acquire it, you can download the model files from the repository using libraries and tools such as huggingface-cli.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-07 | bartowski | GGUF (imatrix) | bartowski/NousResearch_Arete-27B-GGUF | Q2_K 11.8GB (fits in 12GB VRAM) |
| 2026-11-17 | mradermacher | GGUF | mradermacher/Arete-27B-GGUF | Q2_K 12.0GB (fits in 12GB VRAM) |
| 2026-11-17 | mlx-community | MLX | mlx-community/Arete-27B-OptiQ-4bit | MLX 4bit 21.7GB (fits in 24GB VRAM) |
File sizes of each build:
- Available builds in bartowski/NousResearch_Arete-27B-GGUF: IQ2_XXS 8.0GB / IQ2_XS 8.2GB / IQ2_S 8.8GB / IQ2_M 9.6GB / Q2_K 9.9GB / IQ3_XXS 11.3GB / Q3_K_S 11.6GB / IQ3_XS 11.7GB / Q3_K_M 12.3GB / Q3_K_L 12.9GB / IQ3_M 13.6GB / IQ4_XS 14.2GB / Q4_0 15.0GB / Q4_K_S 15.0GB / IQ4_NL 16.0GB / Q4_K_M 16.0GB / Q4_1 16.4GB / Q4_K_L 17.3GB / Q5_K_S 18.0GB / Q5_K_M 19.3GB / Q6_K_S 21.1GB / Q6_K 22.0GB / Q6_K_L 23.0GB / Q8_0 26.7GB / BF16 50.1GB
- Available builds in mradermacher/Arete-27B-GGUF: Q2_K 10.0GB / Q3_K_S 11.2GB / Q3_K_M 12.4GB / Q3_K_L 13.4GB / IQ4_XS 14.2GB / Q4_K_S 14.5GB / Q4_K_M 15.4GB / Q5_K_S 17.4GB / Q5_K_M 17.9GB / Q6_K 20.6GB / Q8_0 26.6GB
- Available builds in mlx-community/Arete-27B-OptiQ-4bit: MLX 4bit 18.1GB
In addition, 37 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Other Models for the Same Task
Recent text generation models covered by Local Model Watch, newest first. Grouped by the task each publisher declares on Hugging Face (pipeline_tag); the smallest VRAM tier is this site’s estimate.
- plamo-3-8b-fin-instruct Text Generation Model: 4GB+ VRAM, GGUF Builds (8.5B, 4GB)
- Qwen3.8-27B-DSpine-b16 Draft Model for Speculative Decoding: 8GB+ VRAM (2.0B, 8GB)
- ByteDance-Seed Releases Cola2 Language Model (8.8B, 48GB)
- EleutherAI Releases Pythia 160M Weight Seed 2 Corrected
See all text generation models →
What to Read Next
- Find models by VRAM (This model runs from the 12GB tier) → Other models that run on a 12GB GPU
- Explore the same model family → Arete family overview (1 articles, 3 converted builds)
- Engines that run this model → llama.cpp / Ollama / vLLM
- What Q2_K, Q3_K_L, Q5_K_M mean and where to get this model → GGUF format guide and models / MLX format guide and models / Quantization and model-format glossary
- Learn about the publisher → Nous Research: models, licenses and articles
Sources
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: mradermacher/Arete-27B-GGUF, mlx-community/Arete-27B-OptiQ-4bit
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of bartowski/NousResearch_Arete-27B-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

