Discrete Flow Matching (DFM) is a novel paradigm designed for generating discrete data, such as language and code. DFM offers several key contributions: it works with a general family of probability paths, allows for a generic formula for sampling using learned posteriors, and improves generative perplexity by focusing on specific probability paths defined with different schedulers. By scaling DFM models to 1.7B parameters, the authors achieve state-of-the-art results on HumanEval and 1-shot MBPP coding benchmarks, demonstrating the capability of generating high-quality discrete data in a non-autoregressive manner. The approach bridges the performance gap between discrete diffusion and autoregressive models, showing significant potential for further enhancements.Discrete Flow Matching (DFM) is a novel paradigm designed for generating discrete data, such as language and code. DFM offers several key contributions: it works with a general family of probability paths, allows for a generic formula for sampling using learned posteriors, and improves generative perplexity by focusing on specific probability paths defined with different schedulers. By scaling DFM models to 1.7B parameters, the authors achieve state-of-the-art results on HumanEval and 1-shot MBPP coding benchmarks, demonstrating the capability of generating high-quality discrete data in a non-autoregressive manner. The approach bridges the performance gap between discrete diffusion and autoregressive models, showing significant potential for further enhancements.