vanta-research/apollo-astralis-4b

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Oct 1, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

Apollo-Astralis V1 4B by VANTA Research is a 4 billion parameter causal language model, fine-tuned from Qwen3-4B-Thinking, designed for conversational reasoning. It uniquely combines rigorous logical thinking with warm, empathetic, and enthusiastic communication, utilizing explicit tags for step-by-step reasoning. This model excels at collaborative problem-solving while maintaining context-appropriate emotional intelligence, making it suitable for applications requiring both analytical depth and engaging interaction.

Loading preview...

Apollo-Astralis V1 4B: A Warm Reasoning Model

Apollo-Astralis V1 4B, developed by VANTA Research, is a 4 billion parameter conversational reasoning model built upon the Qwen3-4B-Thinking base. It is specifically fine-tuned to integrate rigorous logical thinking with a warm, enthusiastic, and empathetic communication style, making it distinct from other LLMs.

Key Capabilities

  • Advanced Reasoning: Employs explicit <think> tags to demonstrate step-by-step reasoning, avoids common logical fallacies, and provides mathematically precise solutions. It also performs critical analysis by questioning assumptions.
  • Warm Communication: Delivers enthusiastic celebrations for achievements, offers empathetic support, and uses a collaborative tone with "we" language and clarifying questions. Its tone adapts to the conversational context.
  • Production-Ready: Maintains a consistent identity, uses natural conversational language, and balances analytical thinking with emotional intelligence.
  • Training: Fine-tuned using LoRA (33M trainable parameters) on a curated dataset emphasizing warmth, empathy, collaboration, and consistent identity.

Good for

  • Collaborative Problem-Solving: Ideal for scenarios requiring both analytical depth and emotionally intelligent interaction.
  • Interactive Assistants: Suitable for applications where a supportive, engaging, and logically sound AI persona is desired.
  • Educational Tools: Can be used in contexts where step-by-step reasoning and clear, encouraging explanations are beneficial.

Limitations

  • Primarily English-focused fine-tuning, despite the base model's multilingual capabilities.
  • Context window of 4096 tokens, inherited from the base model.
  • Not optimized for competition-level mathematics, but strong in conversational reasoning.
  • Enthusiastic style may not suit all professional contexts.

Finetunes
1 models