Stanford and Together AI Preprint Reports HybridAgents Cuts Compute Costs by Up to 73%

A new arXiv preprint authored by researchers including Jiaru Zou of Stanford and Together AI introduced HybridAgents, an AI framework that reduces compute resource usage by roughly 63% to 73% (a factor of 2.7 to 3.7). Evaluated on 4B, 8B, and 14B parameter Qwen3-Base models across 11 math, science, and medicine benchmarks, the framework lowered token usage by 35.5% to 54.7%, accelerated inference speed by 1.8 to 2.6 times, and achieved an average performance gain of 10.4% over existing text-based and latent LLM agents. The architecture relies on a HybridLink module to toggle between internal latent reasoning—which avoids decoding thoughts into text—and standard text mode for external tool calls or final responses, alongside continuous working memory and adapted post-training methods (HybridSFT and HybridRL). While the code and models are publicly available on GitHub and Hugging Face and offer readable pseudo-text for developer inspection, the unreviewed preprint leaves scaling beyond 14B parameters and pseudo-text reliability in regulated domains like medicine as open questions.