Multiverse4FM/Multiverse-32B
Multiverse-32B is a 32.8 billion parameter language model developed by Multiverse4FM, notable for its non-autoregressive architecture. It achieves strong performance in mathematical reasoning, scoring 53.8% on AIME 2024 and 45.8% on AIME 2025. This model is specifically designed for complex problem-solving and advanced reasoning tasks.
Loading preview...
Multiverse-32B: A Non-Autoregressive Model for Advanced Reasoning
Multiverse-32B, developed by Multiverse4FM, is a 32.8 billion parameter language model distinguished by its non-autoregressive architecture. This model represents a significant advancement in tackling complex reasoning challenges, particularly in mathematics.
Key Capabilities & Performance
Multiverse-32B demonstrates exceptional performance in mathematical problem-solving, as evidenced by its scores on challenging benchmarks:
- AIME 2024: Achieved 53.8%, making it the first open-source, non-autoregressive model to surpass 50% on this benchmark.
- AIME 2025: Scored 45.8%, further highlighting its strong mathematical reasoning abilities.
- MATH500: Recorded 91.8% accuracy.
- GPQA-Diamond: Achieved 60.7%.
This model was built upon the Qwen-2.5-32B-Instruct base model and leveraged the s1.1 dataset for training, contributing to its enhanced reasoning capabilities. Its non-autoregressive nature offers a distinct approach compared to traditional LLMs.
Good For
- Advanced Mathematical Reasoning: Ideal for applications requiring high-level problem-solving in mathematics.
- Complex Problem Solving: Suitable for tasks that benefit from a non-autoregressive approach to generate solutions.
- Research in LLM Architectures: Provides a valuable open-source example of a non-autoregressive model for study and development.