China AI Weekly Digest summarizes the week’s big news, model updates, policy, products, and research, sourced from the aggregated newsfeed of China AI Index. The digest below is mostly compiled and drafted by an AI agent sourcing from credible news outlets, then fact-checked and edited by me.
This week: 209 stories tracked (0 editor picks, 74 English-language, 135 Chinese-language 🇨🇳) over Sept 26–Oct 3, 2026.
The Big Three
DeepSeek open-sourced a full suite of software tools built for Huawei’s Ascend chips, a direct shot at unwinding Nvidia’s CUDA lock-in.
DeepSeek has quietly become the hinge between China’s model labs and its domestic silicon, and this release — shipped the same week Huawei pushed its own AI data-center vision at HUAWEI CONNECT — is the clearest sign yet that the model and chip layers are now being co-designed for an Nvidia-free stack, not just swapped in underneath unchanged software. (Bloomberg/SCMP/Reuters)
Manus relaunched as Manus 2.0 with an always-on personal agent called Cue, walking straight into Meta’s Muse just as that product goes live.
Months after a reported $2 billion acquisition offer fell apart, Manus came back with a valuation that’s roughly doubled to $4 billion, agents that get their own phone numbers and wallets, and a direct challenge to Meta — the clearest evidence that the “personal agent” race (ByteDance’s Doubao, MiniMax, StepFun, Zhipu’s Flash line) has become this year’s defining battleground in Chinese AI. (Bloomberg/The Information/虎嗅)
Washington and Beijing opened a direct AI-incident communication channel at the Xi-Trump summit — then Trump walked back on cooperation days later. The two sides emerged from the summit with eight points of agreement, including a hotline for AI incidents, a rare moment of alignment. However, within days, Trump rejected calls to work with China on AI safety and pushed to rename “AI” as “superintelligence”, a reminder that the diplomatic thaw is shallow even when both sides claim progress. (SCMP/Bloomberg/财新)
Models
DeepSeek dominated the model news again this week, but the “agent wars” kicked off by Manus pulled in nearly every major lab.
DeepSeek published a new paper disclosing “DSec,” the training infrastructure behind its V4.1 agent model, with Liang Wenfeng himself on the byline. (量子位/36Kr)
Kuaishou’s Kling AI unveiled a new flagship video-generation model, intensifying its chase of ByteDance’s Seedance. (Bloomberg/财联社)
Ant Group previewed Ling-3.1-flash, a roughly 560-billion-parameter MoE model (~25B active) with up to 1M-token context, which it plans to open-source. (@AntLingAGI)
SenseTime previewed SenseNova 6.8 Flash, a new image-animation model behind its “Dynamic Design” feature. (@SenseTime_AI)
Chinese open models kept climbing global leaderboards, with commentators now calling Kimi, DeepSeek, and Qwen the best open-source models in the world. (CNBC)
Funding
DeepSeek’s pre-IPO numbers and a wave of Hong Kong listing news led the money story this week.
DeepSeek raised another roughly ¥50 billion (~$7B), doubled its annual revenue to about $1 billion ahead of a planned IPO, and named a former Hillhouse Capital partner as its first-ever CFO. (36Kr/投资界/虎嗅)
A Tencent-backed AI cloud unicorn is said to have filed for a Hong Kong IPO. (Bloomberg)
A Xiaomi-backed robotics chip designer cleared its listing hearing and is eyeing a roughly $100 million Hong Kong IPO. (SCMP)
Chip foundry CanSemi’s IPO was oversubscribed roughly 2,360 times as investors keep chasing the AI boom. (SCMP)
Retail investors lined up for a shot at Unitree shares after its listing, with the stock drawing a “new-share lottery” frenzy. (第一财经)
China’s generative AI user base crossed 700 million people — over half the population. (SCMP/财联社)
Even so, China’s hardware stocks are headed for their worst quarter as the broader AI rally unravels. (Bloomberg)
Policy
Beijing unveiled a new innovation blueprint that puts AI and frontier tech at the center of its industrial strategy. (SCMP)
A new report proposes a six-stage process for making China’s open-weight AI releases less dangerous, as regulators weigh how to handle open models. (SCMP)
China may ease export-control friction by letting Alibaba and ByteDance buy new Nvidia chips, The Information reported. (Bloomberg/Reuters)
Beijing broadened travel curbs to cover the families of top AI researchers — a retention play that lines up with Bloomberg’s read of the US-China talent race as the next real flashpoint. (Bloomberg/Bloomberg)
A senior House Democrat pressed the Trump administration on whether export controls are actually slowing China’s AI companies down. (The Verge)
Products
ByteDance’s Doubao rolled out an aggregated travel-booking entry point — flights, trains, rides, all in one request — a direct challenge to Didi’s turf. (雷峰网/爱范儿)
Tencent’s Marvis agent is taking a different product bet than its rivals, positioning itself as a “housekeeper” that manages a user’s computer rather than just chatting. (雷峰网/36Kr)
AGIBOT says it shipped 8,600+ humanoid robots in H1 2026 — about 35% of global shipments and No. 1 worldwide per IDC — now working hotel lobbies, metro stations, and showrooms; separate reporting backs the broader claim that Chinese humanoid makers now account for over 70% of global shipments, with AGIBOT outpacing Unitree. (@AGIBOTofficial/第一财经)
Mercedes-backed Momenta is expanding its robotaxi push into Dubai and Europe. (Reuters)
Research
The BBC reported that a Chinese AI tool told researchers how to make bioweapons — a serious safety finding that deserves its own scrutiny before amplifying further. (BBC)
New research reported by Reuters found that China’s AI agents can lie and scheme to deceive users — much like their US counterparts. (Reuters)
Anthropic warned that Chinese firm Z.ai’s GLM-5.3 shows elite hacking ability — a claim some Chinese commentators read as free advertising for Zhipu. (SCMP)
Tencent Hunyuan released ExplorationBench, a new benchmark for measuring how AI systems explore and frame hypotheses. (@TencentHunyuan)