szili2011/Quanta-X-3B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:3.1BQuant:BF16Ctx Length:32kPublished:Jan 1, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

Quanta-X-3B by szili2011 is a 3 billion parameter Qwen 2.5-based causal language model, enhanced with the Phoenix Framework (DoRA + SimPO Beta 2.0) for deep reasoning. It features an Ouroboros Logic Loop (plan, draft, critique) to ensure logical consistency and reduce hallucinations, making it suitable for complex problem-solving and tasks requiring robust internal thought processes.

Loading preview...

Quanta-X-3B: A Reasoning-Focused Small Language Model

Quanta-X-3B, developed by szili2011, is a 3 billion parameter model built on the Qwen 2.5 base architecture. It distinguishes itself by integrating the Phoenix Framework, which combines DoRA (Weight-Decomposed Low-Rank Adaptation) and an aggressive SimPO (Simulated Preference Optimization Beta 2.0) alignment. This unique training approach is designed to instill "System 2" thinking, enabling deeper reasoning capabilities in a lightweight model.

Key Capabilities & Features

  • Ouroboros Logic Loop: The model employs a hidden reasoning layer where it internally plans, drafts, and critiques its thoughts before generating a final response. This process aims to enhance logical consistency and reduce errors.
  • Diamond-Tier Data Filtering: Quanta-X was trained on a highly curated dataset, including DeepSeek-R1 Traces for logic, OpenR1 Math for verified proofs, Glaive Code V2 for production-ready code, and SlimOrca RP for nuanced storytelling.
  • Hyper-Stability: Through SimPO with a Beta of 2.0, the model was heavily penalized for hallucinations or superficial thinking during training, resulting in a preference for admitting ignorance over providing incorrect information.
  • Extended Context: Supports a 32k context length, utilizing RoPE scaling.
  • Chat Template: Uses the standard Qwen 2.5 ChatML template.

When to Use This Model

  • Complex Reasoning Tasks: Ideal for applications requiring more than instant, superficial responses, where logical consistency and internal validation are crucial.
  • Problem Solving: Suitable for scenarios where the model needs to "think through" a problem, such as logic puzzles or mathematical challenges.
  • Reduced Hallucinations: Beneficial for use cases where factual accuracy and a reluctance to "lie" are paramount, due to its SimPO-driven hyper-stability.
  • Resource-Constrained Environments: Offers advanced reasoning capabilities within a 3 billion parameter footprint, making it efficient for deployment where larger models might be impractical.

Finetunes
1 models