DeepSeek's Platform Play, Qwen's 3 Billion Downloads, and the Manus Unwind — China AI Weekly

DeepSeek moved from model lab to platform company this week — launching V4-Pro to general availability, open-sourcing its Harness agent framework, adopting the AI industry’s first surge pricing model, shipping an experimental vision model, and advancing IPO preparations toward a 2027 listing. Alibaba’s Qwen surpassed 3 billion downloads to become the world’s most-downloaded open-weight model family — more than Meta’s Llama and Google’s Gemini combined. Tencent became the largest shareholder of Manus after China forced Meta to unwind its $2B acquisition — the first time Beijing used its investment security review to reverse a completed AI deal. And China’s first dedicated AI agent governance framework is now operational, even as new cyberspace and data-security regulations take effect. Here is this week’s China AI Weekly.

DeepSeek: The Platform Company Pivot

V4-Pro Goes General Availability

DeepSeek officially moved V4-Pro (build 0813) from preview to general availability on August 13. The flagship model features a 1M-token context window with up to 384K tokens of max output, selectable reasoning effort (low, high, max), and native support for the OpenAI Responses API format.

The architecture is MoE — approximately 1.6 trillion total parameters with ~49 billion active per token. Benchmark scores showed significant gains over the April preview:

On the Artificial Analysis Intelligence Index, V4-Pro scored 53 — up from 45 in the preview, but still trailing Kimi K3 (60), Qwen3.8-Max (58), and Claude Opus 5 (63). Reviews were mixed: strong gains in cybersecurity and coding, but some coverage called non-agentic tasks “underwhelming.”

Notably, weights were not released for this build. The April V4-Pro preview remains the most recent on Hugging Face.

Harness: DeepSeek’s Open-Source Agent Framework

Alongside V4-Pro GA, DeepSeek open-sourced Harness — the internal agent runtime framework it uses for its models. MIT-licensed and positioned as an open alternative to OpenAI Codex and Anthropic Claude Code, Harness lets models directly interact with code repositories: locating files, modifying code, running tests, and iteratively fixing errors.

The response was immediate. Harness accumulated 141,000 GitHub stars in its first 96 hours, making it one of the fastest-growing open-source AI projects in history. On August 21, Harness was updated to v0.1.1-rc.1, adding vision model support and out-of-box integration with DeepSeek’s new V4-Flash-Vision-Exp.

This is a strategic move. By open-sourcing the infrastructure layer, DeepSeek is building a developer ecosystem around its models — not just selling API calls, but creating the tooling that makes those API calls useful. It is the same playbook that made Claude Code and CodEx sticky, but open.

Surge Pricing: The End of the Price War

DeepSeek became the first major AI provider to introduce surge pricing on August 16-17, moving from flat rates to peak/off-peak tiers. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, with peak rates at 2x off-peak.

The steepest increase hit cache hits: from $0.003625/M to $0.022/M off-peak (a 12-fold jump) and $0.044/M at peak. In China domestic pricing, V4-Pro output peaks at 27 yuan per million tokens.

On August 22, DeepSeek announced a weekend discount: Saturday and Sunday will be uniformly billed at off-peak rates, effective August 23.

This is not an isolated move. Morgan Stanley reports that Chinese LLM API prices have risen across the board — average input prices up ~48%, output up ~80% versus Q1 2025. The price war is over. The profit era has begun.

V4-Flash-Vision-Exp: Multimodal Arrives

On August 21, DeepSeek released V4-Flash-Vision-Exp — an experimental multimodal version of V4 Flash that adds image and screenshot analysis. The model matches V4-Flash on text capabilities (agents, reasoning, world knowledge) and approaches Anthropic Opus 4.8 on multimodal agent benchmarks.

Images are tokenized at up to 384 tokens each and billed at V4-Flash pricing. DeepSeek also launched a Files API (free to use — upload once, reference by file_id), and Harness 0.1.1 shipped with native support for the vision model.

The vision model is available via API at model='deepseek-v4-flash-vision-exp'.

Funding and IPO Pipeline

DeepSeek’s second funding round resumed August 6, seeking approximately $8 billion at a valuation near 500 billion yuan (~$74B). The first round, completed earlier in 2026, raised ~$7 billion at a ~$50 billion valuation with investors including Tencent (10B yuan), CATL (5B yuan), NetEase and JD.com (3B yuan each), plus a state AI fund.

The second round requires a minimum 50B yuan per investor. First-round demand was reportedly 2x the 50B yuan raised.

IPO preparations are underway, with filing targeted for 2026 and a potential public debut in 2027. DeepSeek is in discussions with accounting firms and investment banks.

DeepSeek was also a strategic investor in Unitree’s IPO — the humanoid robot company that surged on its market debut. DeepSeek was allocated 933,399 shares at 150.80 yuan, a position worth approximately 788.6M yuan at close — an unrealized gain of ~648M yuan, subject to a 36-month lockup.

Alibaba: Qwen’s Open-Source Dominance

3 Billion Downloads

Alibaba’s Qwen family surpassed 3 billion downloads globally in six months, according to the company. The Hugging Face “State of Open Models: Summer 2026” report (August 14) counted 2,045 million Qwen downloads on the Hub alone — compared to Google’s 418M and Meta’s 227M.

Qwen is now the most-downloaded open-weight model family globally, surpassing Meta’s Llama and Google’s Gemini. The ecosystem includes 460+ distinct open-source models and 300,000+ derivative models (151,448 on Hugging Face alone — 2.6x Meta’s footprint).

A caveat: Hugging Face downloads count pull events, which includes CI pipelines, mirror syncs, and re-downloads that inflate numbers. Alibaba’s 3B figure also includes ModelScope and other platforms beyond Hugging Face. But even adjusting for duplication, Qwen’s lead is substantial.

This is the “great open-source reversal” — the US didn’t see it coming. Two years ago, Meta’s Llama was the undisputed king of open-weight AI. Today, a Chinese model family has more downloads than Meta and Google combined.

Qwen3.8-Max and Qwen3.8-27B

On August 14, Alibaba released two new models:

• Qwen3.8-Max (Qwen3.8-2.4T-A95B): 2.4 trillion total parameters, 95B active (MoE). The first Max-class Qwen with open weights. Artificial Analysis Intelligence Index: 58.
• Qwen3.8-27B: Dense 27B parameter vision-language model, 262K native context (extensible to 1M via YaRN). Apache 2.0 license.

A notable licensing shift: Qwen3.8-Max shipped under a bespoke qwen3.8-max license with non-commercial restrictions and revenue-share requirements — a tightening for a line that leaned hard into permissive Apache 2.0. The 27B variant remains Apache 2.0.

Laptop-ready models also launched, signaling that running frontier AI on consumer hardware is the next battleground.

Office AI Consolidation

Alibaba’s Qianwen Office integrated WeCom on August 18, completing coverage of all three major Chinese office platforms: DingTalk, Feishu, and WeCom (plus Slack for international users). Alibaba also released MyContext as open source — a local context solution for AI agents processing messenger, document, and meeting data.

External models (GLM-5.3, DeepSeek V4 Pro) are now integrated into Qianwen Office’s “Frontier Models” tier. Multi-model is the norm, not the exception.

Reuters Breakingviews: “China’s AI Laggard Could Soon Be Its Leader”

A Reuters Breakingviews piece on August 19 argued that cash-rich Alibaba can lean on e-commerce and cloud profits to fund AI investment, while Tencent draws on gaming and advertising revenue. Baidu’s once-leading Ernie has fallen behind, prompting CEO Robin Li to vow a return to the frontier.

Tencent: The Manus Unwind and WorkBuddy Expansion

The Manus Saga: A First in AI M&A

The most dramatic AI M&A story of 2026 reached its resolution this week.

Meta acquired AI agent startup Manus for $2 billion. China’s NDRC blocked the deal on April 27, citing national security and tech export grounds. Meta unwound the acquisition. Then Tencent led an investor group — with ZhenFund and HSG — to buy back Manus from Meta at the same $2B valuation. Benchmark, previously the largest outside shareholder, did not participate; Tencent bought out Benchmark’s stake.

Tencent is now Manus’s largest shareholder (minority position). Manus resumes independent operations from Singapore.

This is the first publicly disclosed case of China using its foreign investment security review to unwind an already-completed AI acquisition. It signals that Beijing’s regulatory reach extends beyond blocking future deals — it can reverse transactions that have already closed.

Manus’s ARR grew from approximately $100M at the time of Meta’s acquisition to $400-500M during the separation period. As a compliance step, user data generated from December 29, 2025 onward will be deleted on August 23-24.

WorkBuddy Goes Regional

Tencent WorkBuddy v5.0.10 opened WeCom APIs for AI agents on August 18, making 10 core office functions accessible to third-party agents. The model pool includes Hunyuan Hy3, Zhipu GLM, MiniMax, Kimi, and DeepSeek.

WorkBuddy also launched in Thailand on August 15, with the Hy3 model free until August 31. This is Tencent’s first major international push for its AI office product.

Baidu: Ernie Falls Behind

Once China’s undisputed AI leader, Baidu is now playing catch-up. Reuters/Breakingviews reported on August 19 that Baidu’s Ernie model has fallen behind the frontier, and CEO Robin Li vowed to return it to the frontier.

Baidu’s Wenku cloud-storage AI office tool was rebranded to “Kuku AI” on August 14, with a standalone desktop client. Kuku AI claims 25M+ monthly active users for AI office (100M+ total platform MAU) and, like its competitors, now integrates external models including DeepSeek and Zhipu GLM.

Baidu is targeting a Hong Kong primary listing by year-end to fund its AI/ERNIE strategy. Its chip subsidiary, Kunlunxin, is filing separately for a Hong Kong IPO. Ernie Bot passed 200M MAU in January 2026 — impressive user numbers, but the question is whether the model itself can close the capability gap.

Moonshot AI: Kimi K3 and the Compute Loophole

Kimi K3: The “Second Wave” Signal

Moonshot AI’s Kimi K3, released in July, remains the largest open-weight model ever published — a 2.8-trillion-parameter MoE with a 1M-token context window and native vision support. Weights shipped to Hugging Face on July 27.

Kimi K3 scored 57 on the Artificial Analysis Intelligence Index — 4th overall, 1st among open-weight models. It ranks behind Claude Opus 5 (61), Claude Fable 5 (60), and GPT-5.6 Sol (59) overall, but topped the LMArena Frontend Code Arena at 1,679 Elo — beating Fable 5 and GPT-5.6 Sol.

Strong coding and agentic scores: Terminal-Bench 2.1 at 88.3 (self-reported), SWE-bench at 73.8%, Program Bench at 77.8% (#1).

A concern: the hallucination rate rose from 39% (K2.6) to 51%, with accuracy improving from 33% to 46%. The UK AISI/CAISI cyber assessment found Kimi K3 significantly below the most capable US frontier models on cyber, but it outperforms GLM-5.2.

Pricing: $3/M input, $15/M output, $0.30/M cache hit.

The Nvidia Chip Loophole

The White House has accused Moonshot of using Nvidia GB300 chips via a facility in Thailand to train Kimi K3. This is part of a broader pattern: Chinese AI firms — including ByteDance, Alibaba, Tencent, and Moonshot — are accessing banned Nvidia chips through data centers in Thailand, Malaysia, and Japan.

The scheme exploits a gap in US export controls: the regime regulates physical chip ownership, not remote cloud access to compute. Cloud provider Aolani (Singapore-HQ) provides Nvidia chip access in Malaysia, explicitly stating that customers “do not have ownership, potential future claim or physical access to the chips.”

DC Byte data shows 31 planned data centers (100MW+) across Malaysia, Indonesia, and Thailand — versus 2 today. That is a 15x growth pipeline.

The Remote Access Security Act (RASA) passed the House in January 2026 but has stalled in the Senate. Even if passed, enforcement rules would take time. The US-China compute gap is narrowing through multiple paths, and legislation is not keeping up.

Regulatory Roundup: Agent Governance, Data Security, and Cyberspace 2030

AI Agent Governance Framework (Operational)

China’s CAC, NDRC, and MIIT issued the Implementation Opinions on Standardized Application and Innovative Development of Intelligent Agents in May 2026 — the world’s first dedicated AI agent governance framework. It is now operationalizing.

Key provisions:

• Tiered decision authorization: human-only / user-approval-required / agent-autonomous
• Sensitive applications (healthcare, transportation, media, public safety): filing, compliance testing, and product recall mechanisms
• Lifecycle risk assessment and ethics review requirements

This is a first-mover framework. While the EU AI Act addresses AI systems broadly and the US relies on agency-specific guidance, China has moved first on the specific question of autonomous AI agents.

Network Data Security Risk Assessment Measures (Effective August 20)

The national data-risk-assessment regime is now binding. Important-data processors must:

• Conduct annual risk assessments
• Retain reports for 3 years
• Submit within 20 working days when required

Authorities can inspect, require certified assessments, order remediation, or force cessation of important-data processing.

Cyberspace Action Plan 2026-2030 (August 21)

The Central Cyberspace Affairs Commission issued a 5-year plan for cyberspace and information enterprises. Priority areas include:

• High-end chips, basic/industrial software, quantum computing, blockchain, brain-computer interfaces, digital twins
• AI focus: high-end AI chips, large language model performance, and agentic AI technologies
• Open-source participation, IP fast-track channels, enterprise-university-research collaboration
• Global expansion support for competitive firms in computing, cloud, and AI
• Accelerated AI legislation in emerging areas

Filing and Registration Stats

• 538 generative-AI services filed nationally (up 55.5% from 346 in March 2025)
• 263 applications/functions registered (newly disclosed)
• Beijing leads with 242 filings, followed by Jiangsu (80) and Shandong (49)

Hangzhou City-Level Rules (Effective September 1)

Hangzhou becomes the first Chinese city to implement local generative AI service rules. Fines reach up to 5 million yuan for content-safety and IP violations. The rules require algorithm filing before launch and content-safety audit mechanisms.

“AI+” Action Plan and Workforce

The State Council’s “AI+” action plan targets 70% AI penetration by 2027, 90% by 2030. The Ministry of Education is pursuing universal AI literacy by 2030, with AI education across all stages. Provincial policies are following: Shanghai has added AI trainers to Class A vocations (+30% subsidies), and Yangzhou offers 200K yuan training subsidies plus compute vouchers.

Digital Sovereignty Posture

On August 19, Foreign Ministry spokesperson Lin Jian stated China opposes “taking sides” and “forming camps” on AI — a response to reports that the US would ask countries to choose between US and China in the AI race.

A US congressional advisory body reported on August 18 that China is systematically collecting domestic enterprise, operational, and physical-world data that “cannot be scraped” — data crucial for business AI, autonomous vehicles, and humanoid robots. China has elevated data to a core “factor of production” alongside land, labor, capital, and technology.

The AI Office Sector: BAT Consolidation Race

The Chinese AI office sector is consolidating rapidly, framed as a four-way race: DeepSeek (cost-performance), Alibaba (ecosystem), ByteDance (scale), plus Moonshot/Baidu/Tencent chasing.

In the past month:

• Tencent: merged QClaw into WorkBuddy (Jul 20)
• ByteDance: merged Feishu into Doubao (Jul 30)
• Alibaba: unified QoderWork, MuleRun, and Wukong into Qianwen Office (Aug 3)
• Baidu: rebranded GenFlow to Kuku AI (Aug 14)

All platforms now integrate multiple third-party models — DeepSeek, Zhipu GLM, and others. Multi-model is the norm. The question is no longer “which model?” but “which platform integrates them best?”

WorldClaw: The Trump-Linked AI Venture

On August 17, World Liberty Financial — the Trump-backed crypto firm — was reported to be collaborating with WorldClaw, a Hong Kong-based platform offering AI models from Chinese companies that the US has flagged for national security concerns.

43 of the 90 models on WorldClaw come from Alibaba, Baidu, Z.ai, DeepSeek, and Moonshot — companies on the DoD’s Chinese military-aligned list or accused of IP theft. WorldClaw accepts World Liberty’s crypto tokens as payment.

This highlights a fundamental tension: US national security designations restrict American companies from certain Chinese AI providers, but those same models are commercially available through third-party platforms — and now tied to a venture linked to the US president’s family.

The Week in Perspective

Three storylines define this week:

1. DeepSeek’s platform transformation. In roughly ten days, DeepSeek launched V4-Pro GA, open-sourced Harness, introduced surge pricing, shipped a vision model, and advanced IPO preparations. This is not a model lab anymore — it is a platform company with infrastructure, pricing power, and a developer ecosystem. The 141,000 GitHub stars in 96 hours for Harness signal that the developer market is responding.

2. The open-source reversal is real. Qwen alone has more downloads than Meta and Google combined. DeepSeek’s Harness is one of the fastest-growing open-source projects in history. China is not just catching up in open-source AI — it is leading, and by a wide margin. The implications for developer ecosystems, enterprise adoption, and geopolitical influence are significant.

3. Regulation is moving faster than the market expects. China’s AI agent governance framework is the world’s first. Binding data-risk assessments took effect this week. Hangzhou introduces city-level AI rules on September 1. The Cyberspace Action Plan through 2030 explicitly prioritizes agentic AI and legislation for emerging areas. Companies operating in China’s AI space face a regulatory environment that is simultaneously more prescriptive and more supportive than anywhere else.

The throughline: Chinese AI is maturing. From DeepSeek’s surge pricing to Qwen’s licensing tightening to the IPO pipeline, the industry is pivoting from disruption to consolidation, from free to paid, from models to platforms. The question for Western observers is no longer whether China can compete — it is whether the West can keep up with the pace of Chinese AI maturation across models, infrastructure, and governance.

Follow @kkaminsk on X for daily China AI updates. This post is part of the China AI Weekly series.