GenueAI/Tessera-4

Tessera 4: ORPO-Distilled Reasoning Engine

Tessera 4, developed by GenueAI, is a specialized mini-model demonstrating that high-performance reasoning is achievable without massive scale. It leverages ORPO (Odds Ratio Preference Optimization) and distillation from DeepSeek-R1 to excel in logic and mathematics, designed to run efficiently on consumer hardware with 8GB+ VRAM.

Key Capabilities & Performance

• Exceptional Reasoning: Achieves 95% on GSM8K and 93% on ARC-Challenge, surpassing its teacher model, DeepSeek-R1, in these core logic benchmarks.
• Optimized for Logic: Prioritizes logical accuracy over general trivia, with MMLU scores at 66% as a deliberate trade-off for superior reasoning.
• Efficient Training: Developed in approximately 8 hours on a single RTX 3090, focusing on Chain-of-Thought (CoT) path correction to eliminate verbose output.
• Demonstrated Accuracy: Showcases 100% correctness in high-precision math (e.g., 15 factorial), unit conversion, and complex logical branching puzzles.

Hardware & Usage

• VRAM: Recommended 8GB+.
• Compatibility: Optimized for LM Studio, Ollama, and llama.cpp.
• Prompt Format: Requires a specific DeepSeek-V3/R1 style prompt template for optimal performance:

  <|im_start|>system
  You are a highly logical reasoning engine. Think step-by-step.<|im_end|>
  <|im_start|>user
  [Your Question Here]<|im_end|>
  <|im_start|>assistant
  <|thought|>

Good For

• Applications requiring high-precision mathematical and logical problem-solving.
• Scenarios where efficient, accurate reasoning is critical on resource-constrained hardware.
• Developers seeking a model optimized for Chain-of-Thought processing without excessive verbosity.