Tianqi Chen

Tianqi Chen builds open-source machine learning systems spanning models, compilers and hardware. He started XGBoost, Apache TVM and MLC-LLM, and co-created Apache MXNet. His compilation work addresses a practical problem: deploying models across different computing environments without handcrafting every implementation for each device.[2][5]

As of September 18, 2026, Carnegie Mellon University lists Chen as an associate professor in its Machine Learning and Computer Science departments. He leads CMU's Catalyst Group, which studies machine learning systems and AI infrastructure across software and hardware. His personal website also lists him as a distinguished engineer at NVIDIA and records his tenure as OctoAI's chief technologist from 2019 to 2024. CMU announced his NSF CAREER Award for machine learning systems research on June 2, 2026.[3][7][2]

In his August 2023 interview, Chen described MLC's ambition: “The goal is making it deploy universally.” Phone deployments were one illustration. His design separates model development and optimization in Python from deployment into the languages application developers use, including JavaScript, C++ and Java. Compiled tensor functions can then run through the supported language runtime and hardware backend.[1]

Chen explained that compilation can combine operations into fewer GPU kernels, plan memory allocation ahead of execution and transform loops to use hardware efficiently. For language models with dynamic tensor shapes, symbolic shape relationships help the compiler reason about fusion and other decisions even when a dimension's value is unknown. TVM can use existing vendor libraries or generate implementations for platforms where suitable libraries are unavailable.[1]

He also described a deliberate choice to support customization rather than prescribe one quantization method. In 2023, the team was working on “bring your own quantization”: developers would choose the representation while the compiler supplied optimized code. Chen explicitly limited his discussion of int4's memory savings to inference, rather than training.[1]

Chen is a coauthor of WebLLM: A High-Performance In-Browser LLM Inference Engine, first submitted in December 2024. Its April 2026 revision describes a JavaScript framework that runs LLM inference entirely within browsers, with an OpenAI-style API, WebGPU for local GPU acceleration and WebAssembly for CPU computation. MLC-LLM and Apache TVM generate optimized WebGPU kernels, addressing the lack of performant kernel libraries for that environment.[6]

These deployment systems followed Chen's earlier work on tree boosting and deep learning frameworks. His account of that earlier period explains both XGBoost's origins and the manual GPU engineering that preceded TVM.[1]

Chen earned his bachelor's and master's degrees at Shanghai Jiao Tong University, in Yong Yu's ACM Honors Class. His experience list records doctoral study at the University of Washington from 2013 to 2019, advised by Carlos Guestrin. At UW, he also worked closely with Luis Ceze and Arvind Krishnamurthy.[2]

Chen recalled that he and Guestrin initially wanted to test whether alternatives to deep learning could achieve comparable performance when given enough data. Existing gradient boosting libraries were too inefficient for the experiment, so Chen built XGBoost. He later qualified the unsuccessful hypothesis as concerning image classification; he continued to see tree models as useful for tabular data, citing their rule-like decision boundaries, insensitivity to input scale and ability to compose features.[1]

Chen and Guestrin's XGBoost paper, first submitted in March 2016, combined algorithmic and systems contributions: sparsity-aware learning, a weighted quantile sketch for approximate tree learning, and attention to cache access, compression and sharding to make tree boosting scalable.[4]

An unnamed friend persuaded Chen to add Python bindings when he had initially considered a command line interface sufficient. In his recollection, Python bindings, subsequent R bindings and contributions such as visualization broadened the project. He described XGBoost as a community of committers, and MXNet and TVM as substantially more collaborative from their beginnings.[1]

During his master's work, Chen had written CUDA kernels by hand for an ImageNet model, an effort he recalled taking about six months. After XGBoost, collaborative work on MXNet again exposed the difficulty of optimizing models for GPUs and other hardware. That experience prompted his question: “can we build something more generic and automatic?” He connected it directly to starting TVM.[1]

Chen and eleven coauthors presented TVM in a paper first submitted in February 2018. The compiler combines graph-level and operator-level optimization, including operator fusion, mapping to hardware primitives and memory-latency hiding. It uses learning-based cost modeling to explore low-level code optimizations. Chen acknowledged Halide as an early inspiration and explicitly declined credit for inventing machine learning compilation.[5][1]

In an October 2018 technical post, Lianmin Zheng, Eddie Yan and Chen explained AutoTVM's measurement-driven loop: select promising kernel implementations, profile them on real hardware, fit a prediction model to the measurements and use its predictions to choose the next candidates. Their historical comparisons also illustrated a limit of portability: reusing NVIDIA computation and scheduling code without AMD-specific optimization left TVM slower than MIOpen in most tested AMD cases.[8]

Chen sketches high-level architectures in physical notebooks before turning ideas into code. He said this practice began after XGBoost and that the books contain many TVM design details. He described himself as a problem solver willing to work across model design, low-level code and compilers, with enjoyment of coding helping motivate his move into compilation.[1]

Chen advocates open source and open science. He argues that delivering machine learning systems to users exposes real problems and research challenges that prototypes alone may miss. In 2023, he also expressed interest in an ecosystem where personal local models interact with larger cloud models, rather than applications depending entirely on a few large providers.[2][1]

Continuous and lifelong learning are enduring research interests for Chen. He frames them around the whole lifecycle—collecting data, training, fine-tuning and continuously curating local context. His broader view is that useful AI applications require algorithms, systems optimization and data curation to be considered together.[1]