SwarmEvolve

Making math faster: an open-source path to cheaper AI discovery

Google DeepMind's AlphaEvolve shattered a 1969 record by finding faster ways to multiply complex matrices, but it runs on expensive, proprietary Gemini models. SwarmEvolve, a new open-source paper, changes that. It runs swarms of small, latent-reasoning agents on Qwen3 models, reaching scores in five minutes that took other open-source implementations over an hour. This makes cheaper, faster AI-driven math discovery accessible, and we're eager to see it applied to real scientific problems. For more on how LLM inference architecture evolves, explore our related article.