Vaultek/Quartz-R1-8B-Genesis

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Aug 11, 2026Architecture:Transformer Featherless Exclusive Cold

Quartz-R1-8B-Genesis is an 8 billion parameter language model developed by Vaultek, based on the YandexGPT-5-Lite-8B-pretrain architecture. It features an integrated chain-of-thought mechanism using a tag and has been de-censored and fine-tuned using DeepSeek-R1 Distillation and Genesis Tensor Denoising. This model is optimized for reasoning tasks and exhibits high accuracy in stress tests, making it suitable for applications requiring robust analytical capabilities.

Loading preview...

Quartz-R1-8B-Genesis: Integrated Chain-of-Thought Model

Quartz-R1-8B-Genesis is an 8 billion parameter language model developed by Vaultek, built upon the YandexGPT-5-Lite-8B-pretrain architecture. It has undergone significant re-engineering, de-censoring, and fine-tuning using DeepSeek-R1 Distillation and Genesis Tensor Denoising methodologies. A key feature is its integrated chain-of-thought mechanism, requiring the use of a <think> block within its ChatML dialogue template to guide its reasoning process.

Key Capabilities & Features

  • Integrated Chain-of-Thought: Utilizes a mandatory <think> tag for structured reasoning.
  • Genesis Tensor Denoising: Post-training weight filtering, scale alignment, and anomaly removal to enhance precision and reduce hallucinations, even with 4-bit quantization.
  • Robust Performance: Achieves 86.8% on ARC-Challenge and 74.2% on GSM8K, demonstrating strong reasoning and mathematical abilities.
  • Vaultek Custom Stress-Suite: Achieved a 98.0% pass rate on 50 stress tests from a Qwen3.8-27B teacher model and 100% identity accuracy, indicating strong adherence to its intended persona.

Good For

  • Applications requiring explicit, structured reasoning and problem-solving.
  • Tasks benefiting from enhanced precision and reduced hallucinations, particularly in quantized environments.
  • Use cases where a model with a distinct, de-censored persona is desired.