meissosisai/Arc-2
Arc 2 by Meissosis AI is an 8 billion parameter, reasoning-distilled variant of the Llama 3.1 architecture, featuring a 32768-token context length. Engineered to address logical blind spots, it integrates high-density Chain-of-Thought (CoT) data through a 400-step distillation process. This model excels in mathematical reasoning, coding logic, and technical synthesis, making it ideal for applications requiring robust analytical capabilities.
Loading preview...
Arc 2 by Meissosis AI
Arc 2 is an 8 billion parameter, high-intelligence, reasoning-distilled model based on the Llama 3.1 architecture, developed by Meissosis AI. It was specifically engineered to overcome logical "blind spots" common in standard base models by integrating high-density Chain-of-Thought (CoT) data.
Key Capabilities
- Enhanced Logical Consistency: Significantly reduces hallucinations in complex, multi-step word problems.
- Superior Technical Depth: Demonstrates improved performance in Python coding and various STEM subjects.
- Reasoning-Focused Architecture: Utilizes a unique "Anti-Buns" architecture, focusing strictly on high-density reasoning through a 400-step distillation process to balance creative spark with elite logic.
Good For
- Applications requiring strong mathematical reasoning and problem-solving.
- Code generation and debugging, particularly in Python.
- Technical synthesis and complex logical tasks.
- Use cases where reducing hallucinations and improving factual consistency are critical.
Benchmarking Goals (Estimated)
Arc 2 targets significant improvements over Base Llama 3.1 8B:
- GSM8K: 89.5% (vs. 82.0%)
- HumanEval: 76.2% (vs. 67.0%)
- MMLU: 71.4% (vs. 66.0%)
This model was trained using Unsloth on 15 curated datasets, emphasizing reasoning, logic, and knowledge, including open-r1/OpenThoughts-114k-math and bespokelabs/Bespoke-Stratos-17k.