DeepSeek V4 ended a 140-day model silence today — and China’s domestic chip ecosystem was ready before the announcement went live. BAAI’s FlagOS achieved day-zero adaptation across eight chip families simultaneously. Cambricon confirmed native support within hours. The model itself is an efficiency play: V4-Pro uses 27% of V3.2’s inference FLOPs and 10% of its KV cache at 1 million tokens. A 36Kr investigation reveals the five-month delay was caused by a mid-2025 training failure and the decision to migrate training from NVIDIA to Huawei Ascend.
Tencent scrapped its Hunyuan model and rebuilt from scratch in 88 days. Yao Shunyu, recruited from OpenAI in January, led a complete architectural rebuild that produced Hy3 preview — Tencent’s first model under new leadership, with notably improved agent capabilities.
Beijing Auto Show’s opening day revealed two AI platform winners emerging simultaneously. Horizon Robotics launched Starry, China’s first chip integrating cockpit and driving intelligence, while 10+ automakers — BYD, Geely, Li Auto, Great Wall, Changan — simultaneously announced Qwen integration.
China’s government is building formal testing infrastructure for humanoid robots. CAICT launched a “public machine” and “dedicated machine” intake program for its embodied intelligence testing laboratory — the clearest institutional signal that embodied AI is transitioning from demo to deployment. Separately, a CAS-affiliated company claimed the global #1 position on the WorldArena embodied world model benchmark.
OpenClaw crashed hours after reaching WeChat’s 1.3 billion users. Following the enterprise incidents and one-click WeChat deployment we covered this week, the agent framework suffered a major outage — the first large-scale agent infrastructure failure at consumer scale in China.
A one-person AI short drama hit 100 million Douyin views in three days. Production cost for AI short dramas has dropped to roughly 300 RMB per minute versus 30,000 for human-produced equivalents — a 100x cost compression that explains why the format is spreading faster than studios can respond.
DeepSeek released V4 today in two versions: Pro (1.6 trillion total parameters, 49 billion activated, 1 million token context) and Flash (284 billion total, 13 billion activated). The headline isn’t capability — it’s efficiency. V4-Pro uses 27% of V3.2’s inference FLOPs and 10% of its KV cache at 1 million tokens, achieved through a novel DSA sparse attention mechanism. Pricing reflects the thesis: Flash at 0.2 RMB input / 1 RMB output per million tokens, Pro at 1 RMB / 12 RMB. 36Kr, Zhihu
DeepSeek’s internal teams have been using V4 for agentic coding, reporting performance better than Sonnet 4.5 and approaching Opus 4.6 in non-thinking mode. The model remains text-only — no native multimodal support — which 36Kr’s investigative team attributes to compute and cash constraints from the Huawei migration.
That migration is the real story behind the 140-day silence. Insiders told 36Kr that V4’s development was derailed by a serious training failure in mid-2025 and Liang Wenfeng’s decision to migrate the training framework from NVIDIA to Huawei Ascend. Internal disagreements between Liang’s technical ambitions and execution realities slowed progress further. The fundraising window — which we’ve covered since Sunday’s reporting of a valuation escalating past $20 billion — opened specifically because DeepSeek needs capital for a larger next model, and a competitor is reportedly preparing a 3-trillion-parameter model.
Community testing has been mixed. A Zhihu evaluation found V4 performance not clearly superior to GLM 5.1 or Kimi K2.6 — models that shipped while DeepSeek was silent. The 36Kr investigation adds context: Zhipu and MiniMax rushed their latest releases specifically to avoid being overshadowed by V4’s timing, and Tencent’s Yuanbao app is actively trying to reduce DeepSeek dependency (70% of users had DeepSeek as their default model). Tencent bid for 20% equity in DeepSeek and was rejected by Liang Wenfeng.
Jensen Huang warned on the Dwarkesh Podcast this month that “the day DeepSeek launches on Huawei chips first is a terrible consequence for our country.” DeepSeek confirmed Huawei Ascend support for H2 2026 — but the ecosystem around it didn’t wait.
BAAI’s FlagOS — the national AI software stack — completed V4 adaptation across eight domestic chip families simultaneously, achieving what it calls three technical breakthroughs in operator optimization. BAAI Cambricon confirmed native V4 adaptation within hours. Synced DeepSeek itself released TileKernels, an open-source high-performance GPU operator library, alongside V4 — infrastructure designed to make domestic chip deployment easier. Xi Xiaoyao Cloud platforms moved immediately: PPIO announced first-batch V4 hosting and Youdao’s Lobster integrated both V4 and Kimi K2.6 on launch day. 36Kr, 36Kr
This follows the pattern we noted Tuesday — Biren Technology and Huawei Ascend achieving day-zero Kimi K2.6 support — but at a different scale. Eight chip families, simultaneously, on launch day. US export controls intended to slow China’s AI compute have instead produced a coordinated domestic software-hardware stack that demonstrates production-grade readiness whenever a major new model drops.
Yao Shunyu joined Tencent from OpenAI in January and immediately made a call most executives wouldn’t: throw away the existing Hunyuan model entirely and start over. Eighty-eight days later, Hy3 preview went live. GeekPark, Huxiu
Hands-on testing by 36Kr found the results “mixed but improved” — agent capabilities significantly upgraded, general performance still uneven. “Hy3 preview is the first step of the Hunyuan rebuild,” Yao said. This is a foundation, not a finished product.
The speed matters as much as the outcome. Tencent had been a conspicuous laggard in China’s model race — Hunyuan fell behind Qwen, GLM, Kimi, and DeepSeek on most benchmarks. Yao’s willingness to scrap years of accumulated work and bet on a clean architectural restart signals something about Tencent’s internal urgency. Huxiu framed the question: “What’s truly hard to replicate isn’t the model — it’s whether a company has the resolve to rewrite itself for AI.” Hy3 preview launched the same week as DeepSeek V4 and in the aftermath of rapid releases from Qwen, GLM, and Kimi. Tencent is back at the table — arriving as everyone else has already been playing for months.
The 2026 Beijing Auto Show opened with 181 world premieres — and the AI integration announcements revealed a platform layer crystallizing in real time.
On the hardware side, Horizon Robotics, China’s leading automotive AI chip company, launched three products simultaneously: the Starry chip — the first in China to integrate cockpit intelligence and autonomous driving on a single die — the HSD V1.6 full-scenario ADAS system, and KaKaClaw, a vehicle AI operating system with persistent memory that treats the car’s cabin as an agent with personality and skills. Horizon Robotics, TMTPost Two partnership announcements — with Beidou Zhilian and Botek — signaled immediate ecosystem adoption.
On the software side, more than ten major automakers simultaneously announced Qwen integration on opening day: Changan, Dongfeng, BAIC, BYD, Geely, Great Wall, Li Auto, SAIC Volkswagen, and SAIC Zhiji. Alibaba’s Qwen-Omni multimodal model has been adapted for mainstream vehicle chips, enabling in-car route planning, news, hotel booking, food delivery, and package tracking. 36Kr
A Huxiu analysis published the same day revealed the hidden friction: many automakers are internally paralyzed over who within the organization controls the LLM. The question isn’t whether to integrate AI — that debate is over. It’s whether the smart cockpit team, the autonomous driving team, or a new AI department owns the model relationship. Companies that can’t resolve this will default to adopting Qwen or Horizon’s platform — which is exactly how platform winners emerge.
Three developments this week show China’s humanoid robotics sector moving from spectacle to systems.
CAICT formally launched a “public machine” and “dedicated machine” intake program for its embodied intelligence testing laboratory — the first government-backed standardized testing infrastructure for humanoid robots. CAICT When a government standards body builds evaluation rails, procurement frameworks typically follow.
Zhongke Diwuji, a CAS-affiliated embodied AI company, claimed the global #1 overall score on the WorldArena embodied world model benchmark. Robotics Forum UBTECH, one of China’s two leading humanoid robotics companies, released Thinker Cosmos — a developer platform designed to scale the robotics ecosystem beyond its own hardware. UBTECH
In a LateTalk podcast interview, Megvii co-founder Tang Wenbin — who left to start Yuanli Lingji, an embodied AI company — laid out a contrarian bet: skip the humanoid form factor entirely and go modular. His argument: humanoid is overkill for most tasks, and the embodied AI companies that survive will be “a single-digit number, mostly Chinese.” The company is training its own VLM with robot data baked in from the start — not fine-tuning a language model after the fact.
Pony.ai, which already operates robotaxi routes in Singapore and multiple Chinese cities, put a specific number on its cost target for the first time at the Beijing Auto Show: 230,000 RMB (~$32,000) per robotaxi vehicle. Pony.ai
The company also unveiled a joint project with CATL — the world’s largest EV battery maker — producing what they claim is the first L4 autonomous light truck with full vehicle-grade and full-redundancy certification. The fourth-generation autonomous heavy truck entered mass production simultaneously, and Pony.ai announced expanded logistics collaboration with Sinotrans covering both trunk routes and urban delivery. 36Kr
The $32K number matters because it’s the threshold where robotaxi unit economics start working at fleet scale. Waymo’s vehicles cost north of $100K each. If Pony.ai hits its target, the cost advantage becomes structural.
Luo Fuli — one of the four confirmed DeepSeek departures we’ve covered this week, now leading Xiaomi’s MiMo large model team — gave a 3.5-hour interview that amounts to the most detailed public account of where Chinese AI labs believe the next breakthrough lies.
The core claim: Chinese labs have reached internal consensus that the paradigm has shifted from pre-training to post-training and agent reinforcement learning — the path Anthropic pioneered. GPU allocation at top teams has moved to roughly 1:1 between pre-training and post-training, up from heavily pre-training-weighted. The shift represents an “all-in mobilization” across the industry, with organizational restructuring following: flat research tops supported by large-scale human evaluation teams at the bottom.
MiMo-V2.5-Pro, the model Luo Fuli shipped at Xiaomi, entered public beta and is positioning as the top global open-source model on Artificial Analysis. 36Kr She shipped a competitive model before DeepSeek could ship V4 — the talent drain isn’t theoretical.
On the other side of DeepSeek’s departures, LatePost published the first detailed account of why DeepRoute.ai, a Chinese autonomous driving company, recruited Ruan Chong, another core DeepSeek contributor. DeepRoute is pivoting from conventional autonomous driving to large foundation models for driving. LLM talent is flowing into autonomous driving because the two fields now require the same core competency: training large models with reinforcement learning.
A one-person AI short drama called “Anhui Carpenter” hit 100 million views on Douyin in three days. One creator, three days. Jike
The economics are stark: AI short drama production costs roughly 300 RMB per minute; human production runs 30,000 RMB per minute. Following the iQiyi AI actor backlash we covered Monday, TMTPost published an analysis of the widening structural rift between studios pushing AI adoption and creators resisting displacement.
AI virtual idols are already running a near-zero-marginal-cost endorsement flywheel. As one Jike observer noted, every brand paying an AI idol for advertising simultaneously expands the idol’s audience — a loop human celebrities can’t match.
Not everyone is buying in. Pop Mart CEO Wang Ning explicitly rejected AI interactivity for the company’s collectible toys, arguing their value derives from emotional design that AI would destroy. 36Kr His “uselessness is usefulness” philosophy — that products not optimized for utility are the ones people form attachments to — is a deliberate counter-thesis to the AI efficiency drive. Whether Pop Mart’s position holds as AI-native competitors flood the market with VC funding remains open.
OpenClaw suffered a major outage — the first significant failure since ByteDance’s Coze platform enabled one-click deployment to WeChat’s 1.3 billion users earlier this week. XinzhiyuanAI We covered the enterprise-level incidents on Sunday (an infinite API loop at a milk tea chain, a security shutdown at a department store) and the WeChat deployment on Wednesday. The crash is the scale event those incidents were warning about.
Tencent’s WorkBuddy officially launched the same day as an enterprise AI agent platform built on OpenClaw, featuring automatic model routing between cloud and local models. Zhihu WeChat’s agent ecosystem now has both a platform-native deployment path and a reliability problem simultaneously. Agent infrastructure that can handle 1.3 billion users doesn’t exist yet — not in China, not anywhere.
HoYoverse founder building an AI-native game engine: Cai Haoyu, founder of HoYoverse (the studio behind Genshin Impact), is using the company’s LPM 1.0 video model not just for content but as the foundation for a next-generation game development engine. GeekPark
ByteDance’s “model supermarket” bundles top Chinese models for 40 RMB/month: Volcano Engine’s Coding Plan now includes GLM-5.1, MiniMax M2.7, Kimi K2.6, and DeepSeek V3.2 — but developers report aggressive rate limits and hidden deduction multipliers (DeepSeek tokens count as 2x, others up to 5x). 36Kr
CCTV exposes Nezha Auto’s 18.3 billion RMB government fund burn: The once-promising EV brand lost 18.3 billion RMB over three years, with local governments in Yichun and Nanning absorbing billions; 1,631 creditors have filed 26.58 billion RMB in claims against 15.46 million RMB in remaining cash. 36Kr
InnoLight crosses 1 trillion RMB market cap: The optical transceiver maker — a key AI data center component supplier — surged 10x in one year, reflecting capital’s bet on AI infrastructure over applications. Zhihu
Ex-Honor CEO’s ADAS startup targets 8 million vehicles by 2028: Qianli Technology, co-chaired by former Honor CEO Zhao Ming and Megvii founder Yin Qi, announced a StepFun partnership for a driving foundation model, targeting 17x growth from its current 460,000 installed base. 36Kr
Qwen3.6-27B outperforms models 15x its size on coding benchmarks: The compact 27B model beats the much larger Qwen3.5-397B MoE across major coding evaluations — a parameter-efficiency signal for edge deployment. Zhihu
DeepSeek how-to book sells 23 million copies in China: AI instruction manuals have become a publishing phenomenon, signaling mass-market AI tool penetration well beyond the tech community. AIstory
China’s dual-carbon regulation now ties Party promotions to emissions targets: A new assessment framework directly links carbon performance to cadre selection — creating binding constraints on data center expansion and energy-intensive AI compute infrastructure. 36Kr
Funding
Yixing Zhineng raises 1.5B RMB Series B: AI inference chip and infrastructure company, co-led by CICC Capital and Wuxi Venture Capital. ITjuzi
Jiutian Ruixin closes 1.2B RMB Series B++: Compute-in-memory chip startup building an alternative to conventional GPU architectures for inference workloads. ITjuzi
Buzzy raises $20M at $20M ARR: Chinese-founded team (ex-Apple, Google, Adobe, SenseTime) building a natural-language video editing platform positioned as “video Photoshop.” 36Kr