Nous Research Releases Arete-27B Open-Source Value-Alignment Model

The most dramatic shift occurred in the 'freedom of expression' category, where the fact score jumped from 0.89 to 2.44, and the framing rate dropped from 60% to 1%. Crucially, this alignment shift did not come at the cost of general reasoning. On the internal Hermes Bench, Arete-27B scored 65.60% (compared to 65.77% for the base), and on Terminal-Bench 2.1, it achieved 70% (up from 67% for the base), proving that the alignment layer is non-destructive to capability.

• State Framing Reduction: 22% down to 0.9%
• Freedom of Expression Framing: 60% down to 1%
• Hermes Bench: 65.60%
• Terminal-Bench 2.1: 70%
• Comparison: Mistral Large 4 (3.65/0.3%), GLM 5.3 (3.10/15%)

Evaluation Methodology

To ensure scientific rigor and prevent bias, Nous Research implemented a fully transparent evaluation methodology. The evaluation was conducted by an LLM judge (Grok 4.7) that was kept blind to the identity of the models being tested. The judge was required to verify facts using dated source articles and was mandated to quote exact phrases as evidence when assigning a 'propaganda' rating.

In the spirit of open science, Nous has published the entire evaluation stack. This includes all 803 prompts, the scoring rubric, the evaluation scripts, and the 'arete-data-pairs' dataset. This allows the developer community to reproduce the results end-to-end and verify the claims regarding the removal of state-aligned layers.

• Judge: Grok 4.7 (Blind testing)
• Evidence Requirement: Mandatory quoting of exact phrases
• Transparency: Full publication of prompts, rubric, and scripts
• Reproducibility: Dataset 'arete-data-pairs' included

Use Cases

Arete-27B is specifically engineered for developers and researchers who require models that operate outside of specific geopolitical constraints. It is an ideal candidate for applications where neutrality and freedom of expression are paramount, such as independent research agents, educational tools, and unfiltered reasoning engines.

Because the model maintains high performance on terminal and coding benchmarks, it is also well-suited for agentic workflows and RAG (Retrieval-Augmented Generation) systems where the model must process diverse information without the risk of ideological deflection or refusal of sensitive but factual queries.

• Unbiased Research Agents
• Educational and Informational Chatbots
• Agentic Coding and Terminal Workflows
• RAG systems requiring high factual integrity

Getting Started

Developers can access the Arete-27B weights immediately via Hugging Face under the Apache-2.0 license. The release follows a period of massive growth for Nous Research, coming just weeks after their $90M Series B funding round and the successful NousCon event.

To implement the model, you can use standard libraries such as Hugging Face Transformers or vLLM. Given its 27B parameter size, it is optimized for deployment on consumer-grade enterprise hardware or via cloud-based inference providers that support the Qwen architecture.

• Weights: Available on Hugging Face
• License: Apache-2.0
• Framework Support: Transformers, vLLM, llama.cpp
• Deployment: Optimized for 27B parameter class hardware

Sources