Jianyu Huang

Jianyu Huang is a principal research scientist at Meta who builds systems for training and serving large AI models, from low-precision kernels to parallel LLM inference. In 2016, he and his coauthors showed that Strassen's fast matrix multiplication algorithm is practical for small matrices, and he finished his PhD at the University of Texas at Austin under Robert van de Geijn in 2018. After joining Facebook in 2018, he coauthored the FBGEMM quantized-inference library and the DLRM recommendation model in 2019, then in 2021 a system for training recommendation models with up to 12 trillion parameters. In 2024 he was among the coauthors of the Llama 3 model report, and in 2025 he worked on million-token context parallelism and on MetaShuffling for Llama 4 inference. His current work covers reinforcement learning infrastructure and training and inference for frontier models such as Llama 4 and Muse Spark.

Now at: Meta
Works: 21
Badges: Bridge

Publication numbers from OpenAlex, as of 22 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

2012–2015

Microsoft, VMware and Intel Labs

He works at Microsoft Research Asia in Beijing in 2012, at VMware in Palo Alto in 2014 and at Intel Labs in Hillsboro, Oregon, in 2015.
2016

Strassen's algorithm reloaded

With Tyler M. Smith, Greg M. Henry and Robert van de Geijn, he implements Strassen's algorithm with BLIS in 2016 and shows that it is practical for small matrices and for rank-k updates. The work appears at the SC16 supercomputing conference in Salt Lake City in November 2016 under the title "Strassen's Algorithm Reloaded".
2017–2018

A PhD on fast matrix multiplication

He presents families of practical fast matrix multiplication algorithms at IPDPS in 2017 and publishes Strassen's algorithm for tensor contraction in the SIAM Journal on Scientific Computing in 2018. In 2018, he completes his PhD in computer science at the University of Texas at Austin, advised by Robert van de Geijn, with the thesis "Practical Fast Matrix Multiplication Algorithms".
2018–2019

FBGEMM and DLRM at Facebook

He joins Facebook in Menlo Park in 2018. In 2019, he coauthors FBGEMM, a low-precision kernel library for quantized inference that delivers more than 2x performance gains over Facebook's production baseline, and the Deep Learning Recommendation Model (DLRM), released with implementations in PyTorch and Caffe2.
2021–2022

Training 12-trillion-parameter recommenders

In April 2021, he coauthors a software-hardware co-design for training deep learning recommendation models at Facebook, pairing a PyTorch-based stack with the ZionEX platform. The system trains models with up to 12 trillion parameters at a 40x speedup in time to solution, and the paper appears at ISCA in June 2022.
July 2024

The Llama 3 herd of models

In July 2024, he is among the coauthors of "The Llama 3 Herd of Models", whose largest model is a dense Transformer with 405 billion parameters and a context window of up to 128K tokens. He also works on FBGEMM GenAI, PyTorch GPU operator libraries whose FP8 quantization kernels are used in Llama 3.
2025

Million-token context parallelism

At MLSys 2025 in Santa Clara, he and colleagues present context parallelism for long-context inference, reaching a 1M-token prefill with Llama 3 405B in 77 seconds. In May 2025, he coauthors MetaShuffling, a mixture-of-experts inference method for deploying Llama 4 models in production, and in October 2025 he co-writes a Meta engineering post on scaling LLM inference for the Meta AI App.
As of October 2026

Principal research scientist at Meta

He is a principal research scientist at Meta Platforms in Menlo Park, working on reinforcement learning infrastructure and high-performance training and inference for frontier LLMs, including Llama 3, Llama 4 and Muse Spark. In 2026, he coauthors an MLSys 2026 paper on optimizing deployment configurations for LLM inference.