AI · Robotics · Quantum · Deep Tech

Decoding China's Hard Tech Frontier.

Your premier source for China's breakthroughs in AI, robotics, and deep tech. We translate arXiv papers and track top open-source projects to bridge the global innovation gap.

120+Papers analyzed
350+OSS projects tracked
5,000+Weekly readers
100%Independent & free
Tech Radar

Tech Radar: What's Trending in China

Real-time updates on China's deep tech ecosystem.

AI

Tsinghua's Agent Framework Passes Human-Level Multi-Step Planning Benchmark

New open framework chains tool use, memory, and self-reflection — scoring 91.2 on a benchmark previously dominated by frontier Western labs.

Tsinghua Univ.2 hours ago
EV

CATL Claims 1,500 km Solid-State Battery Enters Pilot Production

The condensed-matter cell promises 500 Wh/kg density and 10-minute fast charging; mass production targeted for next year's flagship sedans.

CATL / Fujian6 hours ago
Quantum

USTI Photonic Quantum Computer Solves Gaussian Boson Sampling 10⁶× Faster

Upgraded Jiuzhang 4.0 pushes photonic qubit count to 510, cementing China's lead in optical quantum advantage experiments.

USTC HefeiYesterday
Chips

SMIC Reportedly Tapes Out 3nm-Class SoC on Domestic DUV Multi-Patterning

Analysts flag a major yield milestone for homegrown lithography toolchains — a pivotal step toward semiconductor self-sufficiency.

SMIC Shanghai2 days ago
Robotics

Unitree's Humanoid Now Runs Tesla-Factory-Style Shifts at Two Auto Plants

G1-based workers complete 8-hour battery-swap cycles in pilot lines, with unit costs reportedly below a US$30k threshold.

Unitree Hangzhou2 days ago
Space

Commercial Reusable Rocket Completes First Vertical Landing Test in Haiyan

The 65m kerosene launcher nailed a 100m hop test — China's answer to Falcon 9 reusability edges closer to orbital debut.

LandSpace3 days ago

No stories match your search — try another keyword.

Open Source Hub

Open Source Hub: China's Code Going Global

Top GitHub repositories and open-source tools from Chinese developers.

DeepSeek-Infer

Blazing-fast MoE inference engine powering the DeepSeek model family — 3× throughput over vanilla vLLM on consumer GPUs.

98.2k Stars 9.1k Forks Updated 1d ago
View on GitHub

AgentScope-X

Multi-agent orchestration framework from Alibaba DAMO — drag-and-drop graph UI, 40+ built-in tools, native distributed execution.

41.5k Stars 4.8k Forks Updated 3h ago
View on GitHub

RoboVerse

Unified simulation & real-world training stack for humanoids from Shanghai AI Lab — photorealistic sim2real in one config file.

23.7k Stars 2.9k Forks Updated 8h ago
View on GitHub

YOLO-Next

The community successor to YOLOv8 maintained in Wuhan — real-time detection at 190 FPS on Jetson Orin, with edge-quantized exports.

31.9k Stars 6.2k Forks Updated 12h ago
View on GitHub

QuLeaf SDK

Full-stack quantum programming SDK from Baidu Research — circuit design, transpilation, and cloud access to superconducting QPUs.

12.4k Stars 1.6k Forks Updated 2d ago
View on GitHub

FireFly-Tune

One-click distributed fine-tuning for 100B+ models on a single 8-GPU node — from the team behind China's most-starred LLM cookbook.

28.1k Stars 3.3k Forks Updated 5h ago
View on GitHub

No projects match your search — try another keyword.

Paper Insights

Paper Insights: Lab to Reality

Deep dives into the latest arXiv papers from top Chinese AI labs.

Acc@len → 96.4%
LLM

[Paper] Scalable Chain-of-Thought via Recursive Depth Verification

《基于递归深度验证的可扩展思维链推理》

  • Introduces R-DPV, a recursive verification loop that cuts reasoning errors by 43% on GSM-Hard without any extra training data.
  • Accuracy keeps climbing with test-time compute — the first open recipe to cleanly scale CoT depth past 50 steps.
  • Runs on a single A100: full training code and 32k-sample curriculum released under MIT license.
arXiv:2601.04721 · Tsinghua KEG Lab · Jan 2026
grasp success: 97.1%
Robotics

[Paper] DexGraspNet-2: Language-Conditioned Dexterous Manipulation at Scale

《DexGraspNet-2:语言条件驱动的大规模灵巧手操作》

  • A 1.2M-pair language–grasp dataset generated in simulation, validated on real 3-finger and 5-finger hands with 97.1% success.
  • Zero-shot transfer across 40 novel household objects — no per-object calibration needed.
  • Sets a new SOTA on the DexBench suite, beating prior methods by 18 points on unseen categories.
arXiv:2601.03882 · Shanghai AI Lab · Jan 2026
spatiotemporal tokens
Computer Vision

[Paper] WorldDreamer-2: Long-Horizon Video World Models with Memory Tokens

《WorldDreamer-2:基于记忆令牌的长时程视频世界模型》

  • Generates coherent 60-second video with persistent object memory — scenes stay consistent across camera cuts and occlusions.
  • Memory-token compression reduces inference cost 5.8× vs. naive frame-conditioned diffusion.
  • Doubles as a neural simulator for autonomous driving: closed-loop policy training transfers to CARLA with 91% fidelity.
arXiv:2512.19407 · Zhipu AI / Tsinghua · Dec 2025

Stay Ahead of the Curve.

Join 5,000+ global tech enthusiasts. Get our weekly digest of China's top open-source projects and paper breakdowns straight to your inbox.