Research Intern, Frontier Speculative Decoding (Winter 2027)

<h3><span style="font-family: helvetica, arial, sans-serif;"><strong>About the Role</strong></span></h3> <p><span style="font-family: helvetica, arial, sans-serif;">The Speculative Decoding team at Together AI builds the draft models and algorithms that let large language models generate faster without changing their outputs. Our speculators run in production on frontier open models such as DeepSeek, Kimi and GLM. Our work sits at the intersection of model training, inference algorithms and serving systems, and we co-design all three to lower the cost and latency of generation.</span></p> <p><span style="font-family: helvetica, arial, sans-serif;">As a research intern, you will dive into open questions in speculative decoding: new drafter architectures and verification schemes, knowledge distillation and training recipes that align draft and target models, and drafters that adapt as traffic shifts. You will train and fine-tune draft models at scale on dedicated GPU clusters, and test your ideas against realistic serving workloads, including long-context and agentic traffic.</span></p> <p><span style="font-family: helvetica, arial, sans-serif;">Projects aim to answer important research questions in speculative decoding, enabling faster generation on frontier models, drafters that keep pace with changing workloads, and robust, reproducible evaluation under real traffic.</span></p> <h3 id="past-research" data-lfile-container="75b2226e-f583-4bc8-a073-fba516e11feb" data-pm-slice="1 3 []"><span style="font-family: helvetica, arial, sans-serif;">Past Research</span></h3> <p><span style="font-family: helvetica, arial, sans-serif;">Speculative decoding research by Together researchers, interns and academic collaborators includes:</span></p> <ul> <li style="font-family: helvetica, arial, sans-serif;"><span style="font-family: helvetica, arial, sans-serif;"><a href="https://openreview.net/forum?id=aL1Wnml9Ef" target="_blank">Speculative Speculative Decoding</a> (ICLR 2026)</sp