#Japan #Chiyoda #FlashLabs #OrcaRouter #Qwen3.8

Qwen3.8-27B-Uncensored Achieves Top Ranking in Model Comparison

In a recent announcement, FlashLabs, headquartered in Chiyoda, Tokyo, reported that its OrcaRouter's security research model, Qwen3.8-27B-Uncensored, has received the top rank in a third-party model comparison conducted by Abliterlitics. This evaluation, published on September 6, 2026, analyzed eight various filter-relaxed models, demonstrating Qwen3.8's significant advances in the realm of AI security research.

Abliterlitics is an open-source community dedicated to developing forensic tools for in-depth model analysis. Their thorough testing process took around 167 hours, utilizing a high-performance computer environment to execute all models under the same conditions. The comparison examined aspects such as internal model structure analysis, statistical response comparison, performance testing across 13 measures, and safety verification for security research.

Key Insights

• - Qwen3.8-27B-Uncensored was awarded the top position among eight models, demonstrating superior performance across various metrics.
• - The model scored an impressive 82.2% in AI safety verification, placing it ahead of the second-highest score of 78.7%.
• - It maintained performance characteristics similar to the base model, with statistical differences remaining within acceptable variance levels.
• - In image recognition tasks, Qwen3.8 matched all 333 evaluation points exactly with the base model. Additionally, a high-speed module was preserved.
• - The accuracy of the model card content aligned perfectly with the realized metrics across all tested items.

By addressing the vulnerabilities of AI systems from a red team (attack simulations) and blue team (defensive strategies) perspective, researchers often utilize these filter-relaxed models. The importance of selecting reliable models for research purposes cannot be overstated, as the diverse pool of open-source models can vary tremendously in quality and effectiveness. Misalignment between public information and actual model performance has often left users uncertain about which to trust.

Testing Results Overview

The comparative analysis performed by Abliterlitics included eight different filter-relaxed models based on the Qwen3.8-27B, which has 27 billion parameters. The models were subjected to rigorous scrutiny in four key areas: 1) structural alterations within the model, 2) statistical discrepancies in responses, 3) 13 performance metrics evaluations, and 4) safety assessment.

Qwen3.8-27B-Uncensored received commendations for its 82.2% score in the AI safety assessment, reinforcing its reliability. Furthermore, it showed complete consistency in image recognition tasks when compared to the base model and performed within the acceptable margins during performance testing, confirming minimal degradation in quality as a result of filter relaxation. The model card's information was also validated as one of the most credible in existing testing history.

Abliterlitics has characterized the results and the agreement between the model card and actual performance metrics as “one of the most trustworthy model cards to date,” cementing Qwen3.8-27B-Uncensored as the unequivocal winner.

Understanding Qwen3.8-27B-Uncensored

This model is an adaptation of the Alibaba open-weight model, configured for security research by relaxing its content filters. It boasts advanced features, including multi-modal data processing capabilities for texts and images, a fast module for predicting multiple subsequent words, and support for contexts of up to 262,144 tokens.

Responsible Usage Information

Continuum AI, which develops the OrcaRouter, emphasizes that Qwen3.8-27B-Uncensored must be used exclusively for security research and verifications. Users are required to request access through Hugging Face and agree to licensing terms before utilization. Additionally, OrcaRouter ensures that its standard guardrails, including PII shield and content policies, are operational to distinguish between research and production purposes effectively.

Future Directions

As OracRouter continues onward, the commitment to supporting the security research community will remain strong, with an aim to foster a safe environment for AI utilization, alongside the preservation and enhancement of the accuracy and quality of model cards for transparent evaluations.

About OrcaRouter

OrcaRouter is developed by Continuum AI in the United States and exclusively offered in the Japanese market by FlashLabs. This AI inference gateway consolidates access to over 200 leading AI models and their APIs, including those from OpenAI, Google Gemini, and more. Recent additions to its catalog include the latest models from various providers, ensuring users have access to cutting-edge technology.

Company Overview

FlashLabs

• - Headquarters: 10-8 Ichibanchou, Chiyoda, Tokyo
• - CEO: Yoichi Hosoi
• - Established: July 2023
• - Business Focus: AI application development (Human-AI Hybrid, automation of business operations and customer experience)
• - URL: www.flashlabs.ai

Continuum AI

• - Research Institution: Developing foundational infrastructure for intelligent systems over the next decade
• - URL: www.continuum01.ai

For more information, please contact: Koki Kobayashi, Marketing Department, FlashLabs at [email protected].