For years, one number decided whether a large model could be trained at all: how fast the optimizer could chew through a batch. The person who moved that number twice — from fourteen days to minutes — is now running a company, and his next argument is that the whole stack underneath today's models is about to be rebuilt from scratch.
Key takeaways
- Yang You (尤洋), a Presidential Young Professor at the National University of Singapore, wrote the LARS and LAMB optimizers; his ICLR 2020 paper cut BERT pretraining from three days to 76 minutes.
- He co-founded Beijing-based Luchen Tech (潞晨科技) in July 2021 and built Colossal-AI, an open-source distributed training system whose paper lists him as a co-author.
- Colossal-AI had more than 41,000 GitHub stars and Open-Sora more than 29,000 when checked for this profile; Luchen Tech raised a Series B of several hundred million yuan in March 2026.
- Luchen Tech reported 2025 contracted revenue of 260 million yuan (≈ US$37 million / HK$283 million) with overseas revenue at 48% of the total.
- You won a 2026 Google Research Award to build systems foundations for diffusion-based large language models, on the argument that today's inference stack is built for sequential generation and largely incompatible with diffusion.
Two papers that moved the field
You's academic arc is unusually compact. He did his undergraduate and master's work at Tsinghua University (清华大学), then a doctorate at UC Berkeley under James Demmel, and joined NUS's School of Computing as an assistant professor in August 2020 — the department's own announcement notes he arrived with the Lotfi A. Zadeh Prize and a Siebel Scholar award. His faculty page lists an IPDPS 2015 best paper, an ICPP 2018 best paper, the ACM/IEEE George Michael Memorial HPC Fellowship, an ACM doctoral dissertation award nomination from Berkeley, and a Forbes 30 Under 30 Asia listing in 2021.
The two papers are the durable part. LARS, in "ImageNet Training in Minutes," showed that a layer-wise adaptive learning rate could hold accuracy at very large batch sizes, finishing a 100-epoch AlexNet run in 11 minutes on 1,024 CPUs and a 90-epoch ResNet-50 run in 20 minutes on 2,048 Intel Knights Landing processors without losing accuracy. LAMB, in "Large Batch Optimization for Deep Learning: Training BERT in 76 minutes," found that LARS did not transfer to attention models, and proposed a layer-wise strategy that allowed a batch size of 32,768 with no degradation, cutting BERT pretraining from three days to 76 minutes on a TPUv3 pod. The NUS faculty page notes that LARS and LAMB are both included in the MLPerf industry benchmark suite.
The pattern is worth stating plainly: You kept finding that the binding constraint was not the model but the plumbing around it.
From a paper to a company
In July 2021, You founded Luchen Tech in Beijing. The product built around his research, Colossal-AI, is a unified system for distributed training that supports data, pipeline, tensor and sequence parallelism plus heterogeneous training; the accompanying paper reports up to 2.76 times training speedup against its baseline. Open-Sora, an open video-generation model, followed in 2024.
The open-source footprint became the funnel. Checking GitHub directly for this profile, Colossal-AI showed more than 41,400 stars and Open-Sora more than 29,900 — together a distribution footprint that most venture-backed infrastructure companies pay for. Luchen Tech also published a further open-source tool, Open-Sora, and gained the attention of investors including Sequoia China (红杉中国), Innovation Works (创新工场), ZhenFund (真格基金), Prosperity7 Ventures, C Capital, Alibaba and Sea Limited.
In March 2026 the company announced a Series B of several hundred million yuan, led by Bohua Capital (博华资本) jointly with an unnamed US-dollar fund, earmarked for cloud productisation. Alongside it, Luchen Tech disclosed commercial numbers: contracted revenue of 260 million yuan for 2025 with 120 million yuan delivered and collected, cloud and software at 75% of revenue and overseas at 48%. By the first quarter of 2026, cloud product revenue (PaaS and MaaS) had reached 87% of the mix and overseas revenue 79%.
That last pair of numbers is the interesting part. A systems company that began as an academic open-source project now earns most of its money outside China, and most of it from products rather than projects.
The next argument: diffusion needs its own stack
You's research direction has moved again. He now works on diffusion-based large language models, which refine all tokens in parallel instead of emitting them one at a time. In the 2026 Google Research Award announcement published by NUS Computing, he argues that the infrastructure behind current models was built over a decade for the sequential approach and mostly does not work for diffusion: there is no diffusion-native optimiser, no equivalent of FlashAttention for bidirectional attention, and KV caching is fundamentally incompatible.
His team's answer is a set of foundations rather than a single model — AsyncLane for block-diffusion inference, ECUpcycle for converting a dense autoregressive model into a diffusion mixture-of-experts model, and OrScale, an optimiser designed for diffusion training. The group's first multimodal diffusion work, DiffuSpeech, is already released. You has been consistent about method: build the systems layer first, then the model capability follows.
Honest limitations
This profile draws on primary sources: the NUS School of Computing faculty directory and department announcements, the arXiv records for the LARS, LAMB and Colossal-AI papers, the ACL Anthology entry for the CAME optimiser, and GitHub's API for star counts checked on the day of writing. The Luchen Tech revenue, mix and funding figures come from the company's own announcement as reported by Daily Economic News (每日经济新闻) and Tencent News; they are unaudited company statements and I have not independently verified them. Star counts move daily and are stated here as a snapshot. I deliberately excluded a 2025 public dispute over model cost estimates: it is not relevant to his technical record, and I could not verify the original documents to describe it fairly, so it is omitted rather than summarised. This profile also does not attempt to assess the commercial claims behind Colossal-AI's benchmark numbers, which are vendor-reported.
What readers can do now
- Read the LAMB and Colossal-AI papers directly if you maintain training infrastructure; both describe concrete techniques rather than slogans.
- Treat GitHub star counts as a distribution signal, not a quality signal, and benchmark the framework on your own hardware and model sizes.
- If you serve models, the diffusion-argument section is worth tracking: You's team predicts that serving costs and response latency change shape if diffusion-based language models become dominant.