🏆 Championship at #MLSys2026! Huge congrats to @Tsinghua Storage Research Group & Tencent for winning the MoE Kernel Challenge, dropping Qwen3-30B-A3B inference latency from 14.91s to 3.56s using NKI on AWS Trainium2/3, delivering a whopping 4.12x speedup. 🚀 Outstanding work!
With AI9Stars, Innovative Tsinghua team @TsinghuaNLP introduces MERLIN, a signal-text multimodal model with a new dataset and benchmark. It uses feature-level denoising to stay robust in noisy environments and enable reliable electromagnetic signal understanding and reasoning.
One model for both seeing and creating! Innovative Tsinghua team @TsinghuaNLP presents CHEERS, an open-source end-to-end framework that unifies multimodal understanding and generation with joint optimization, and strong performance with less data & tokens.
Excited to see this reflection from the German delegation after their visit to our Center for Artificial Intelligence General Education (AIGE) at Tsinghua DCST. Glad that our work on AI literacy as a general education resonates globally. 🌐https://t.co/dL8cqvNYHP
#Tsinghua DCST handball team captured #MaYuehanCup championship, completing a stunning comeback to defeat the three-time defending champions 26:22. Led by Pavlov’s explosive offense and the whole team’s relentless defense, we finally stand at the top. 💪🤾
The 12th NExT++ Workshop recently wrapped in Singapore! 🌏Experts from @Tsinghua_Uni, @NUSingapore and @unisouthampton gathered for in-depth discussion on #ResponsibleAI. From #AgenticAI to global governance, we continue to drive global collaboration in trustworthy AI.
As part of the #TsinghuaFamily, the Department of Computer Science and Technology embraces #DiversityAtTsinghua for #Tsinghua115! We’ve just moved into our new Baichuan Building. Alumni, come home to explore our fresh space and relive old memories. Welcome back!
Running LLMs 5× faster without extra GPU memory? Innovative Tsinghua’s Prof. Liu Zhiyuan team @thudcst@TsinghuaNLP introduces NOSA, a trainable sparse attention that efficiently offloads KV caches while keeping CPU-GPU transfers low.
Long-context AI without massive retraining? Innovative Tsinghua introduces HALO & HypeNet, a hybrid Transformer–RNN delivering comparable performance with superior efficiency using just 2.3B tokens, <0.01% of pre-training data @TsinghuaNLP.
@YiyangDu21@Weixin_WeChat To advance the field of AI, Du pursued research at @AIRTHU1201, exploring multimodal LLMs, later continuing this work at @Stanford through a summer research program. His work was eventually accepted by leading conferences, including @aclmeeting & @CVPRConf.
@YiyangDu21 applied his technical skills by developing a @Weixin_WeChat mini program for class activities. He later joined DeepLang AI, where his team won the championship in the Beijing division of the China College Students’ “Internet+” Innovation & Entrepreneurship Competition
“Computer science is not just textbook knowledge, but a tool to change the world,” says TsinghuaRen Du Yiyang (China), an alumnus of @thudcst. Guided by curiosity and passion, he ventures into the “unknown” of abstract code to create real-world solutions that benefit people.
#Tsinghua DCST shuttlecock team just made history by winning the #MaYuehanCup Group A title and claiming the team's first-ever overall university championship trophy since its founding!🥇 Huge props to Gu Yuxian, Zheng Kehan, You Wenxun, and Tang Songqi for an incredible run.🏆
After just 3 years since its founding, #Tsinghua DCST Fencing Team has won the #MaYuehanCup Group A Epee Team Champion!
Not only that:
Perevoznikova Alisa (Alice) – Women’s Epee Champion
Lu Yan – Men’s Foil Top 4
Yuan Zerui – Men’s Epee Top 8
Proud of you all! ⚔️
Congrats to alumnus Laurence Tianruo Yang @thudcst !🌟 He received the 2026 @ComputerSociety Edward J. McCluskey Technical Achievement Award for pioneering Cyber-Physical-Social system design and data analytics with global impact.
Congrats to #Tsinghua DCST's Storage Research Group for winning the Distinguished Artifact Award at #USENIX#FAST26 for their paper "GPU Checkpoint/Restore Made Fast and Lightweight", slashing checkpoint times by up to 72% with <1% overhead. Boosting GPU efficiency! 🚀
Proud moment for #Tsinghua DCST! Our 2 PhD students just won a Distinguished Paper Award at #NDSS2026 for "OCCUPY+PROBE" – a groundbreaking BTB side-channel attack on Intel CPUs that extracts RSA keys & breaks KASLR. A huge step forward in #CyberSecurity. 🚀 #NDSS
AI for justice! At CJAI2026, Innovative Tsinghua @thuir_lab, @thudcst and collaborators released LegalOne-R1, a 8B model achieving professional-level performance comparable to DeepSeek-R1 and GPT-5 on Chinese legal tasks, supporting legal tech in courts.
“As models & data sets grow, emergent abilities can appear. This non-linear uncertainty is what makes large models fascinating,” said Prof Sun Maosong @thudcst. However, he claims most AI-related firms will benefit more from chasing precise uses over size.
"The relationship between AI and doctors is symbiotic, not substitutive," emphasized Prof. Zhang Bo. Future physicians must master AI to amplify their humanistic care, clinical judgment, and collaborative expertise.