China AI Weekly: DeepSeek V4-Flash's Agentic Leap, Qwen 3.8 Max's 2.4T Multimodal Giant, and China's First Binding AI Agent Rules

The week of July 24 – August 1, 2026 was a blockbuster for Chinese AI. DeepSeek shipped the official V4-Flash API with agentic benchmark scores that beat its own larger sibling. Alibaba previewed a 2.4-trillion-parameter multimodal model. Moonshot open-sourced Kimi K3 — the largest open-weight model in history — and had to pause sign-ups 48 hours later because demand literally broke their infrastructure. MiniMax launched an H3 video model. China’s first binding AI agent regulations took effect on July 15. The government declared that refusing to use domestic chips is “treason.” And Chinese models now hold six of the global top 10 leaderboard spots, winning on price by 90–99% discounts against American counterparts.

This is not convergence. This is a structural realignment of the global AI landscape, happening in real time.

Here is our professional analysis.

DeepSeek: V4-Flash, V4-Pro Delays, and an IPO Path

V4-Flash: The Small Model That Beat the Big One

On July 31, DeepSeek released the official DeepSeek-V4-Flash-0731 API in public beta. The architecture is the same 284B total / 13B active Mixture-of-Experts design from the April preview — but the performance gains are driven entirely by extensive post-training, not parameter scaling.

The benchmark results are striking:

• Terminal-Bench 2.1: 82.7% — up from 61.8% in the April preview, and beating V4-Pro-Preview’s 72.1%. The smaller model outperforms the larger one on agentic tasks. This is the post-training scaling thesis in action: that inference-time and post-training techniques can substitute for raw parameter count.
• Artificial Analysis Intelligence Index: 50 — a 10-point jump, now 6 points ahead of V4-Pro and within 1 point of GPT-5.6 Luna (51) and GLM-5.2 (51).
• Native support for the Responses API format and Codex compatibility — making migration from OpenAI-based workflows straightforward.

The pricing continues DeepSeek’s relentless cost compression: $0.14 per million input tokens, $0.28 per million output tokens. Cache hits drop to $0.0028/1M — a 98% discount. For context, Anthropic’s Fable 5 costs $50 per million output tokens. DeepSeek is offering frontier-adjacent performance at roughly 0.6% of Anthropic’s pricing.

MIT open weights are coming to HuggingFace. Legacy deepseek-chat and deepseek-reasoner endpoints are deprecated, with a migration deadline of October 24, 2026.

The strategic signal is clear: DeepSeek is betting that post-training quality, not scale, will define the next frontier of AI value. If V4-Flash at 13B active parameters can match models 3–4x its size on agentic benchmarks, the parameter race may be giving way to a post-training race — where DeepSeek has a structural advantage in research talent and training methodology.

V4-Pro GA Pushed to Mid-August

V4-Pro (1.6T total params, 49B active) remains in preview, with general availability now reported for August 10–20, pending closed testing of DeepSeek’s in-house “Harness” coding tool. V4-Pro-Max scores 80.6% on SWE-bench Verified — elite territory for autonomous coding. The delay suggests DeepSeek is ensuring the GA release meets production-grade reliability for enterprise coding workflows, not just benchmark milestones.

V4-Pro pricing: $0.435/$0.87 per million tokens. Still dramatically cheaper than Western counterparts, but notably more expensive than V4-Flash — suggesting DeepSeek sees a two-tier market: Flash for high-volume, cost-sensitive workloads, and Pro for premium enterprise and coding use cases.

Funding Round Halt and IPO Path

DeepSeek suspended its second funding round (target: ≥10B RMB) on July 26, after a transcript of founder Liang Wenfeng’s closed-door May investor meeting leaked to Chinese media. The 3 hour 44 minute transcript reportedly discussed technical lag and dependence on foreign compute — politically explosive topics given Beijing’s domestic-chip mandate.

The first round closed in June: $7.5 billion from Tencent, CATL, and the National AI Investment Fund. DeepSeek is reportedly preparing for an IPO on the Shanghai STAR Market as early as Q2 2027, targeting a valuation up to $71 billion.

The IPO path is a structural necessity. China’s private capital pool is shallower than Silicon Valley’s, and frontier AI burn rates cannot be sustained through private rounds indefinitely. But public markets demand quarterly results and revenue growth. How DeepSeek navigates the collision between its AGI-first research culture and market discipline will be one of the defining corporate governance stories of 2026–2027.

DSpark: 85% Inference Speedup

DeepSeek unveiled “DSpark,” a speculative decoding framework that speeds up inference by up to 85%, easing bottlenecks and chip strain under US export controls. This is not just a performance optimization — it is a strategic response to compute constraints. If you cannot buy more chips, you make each chip do more. Speculative decoding lets a smaller draft model propose tokens that a larger model verifies, dramatically reducing the compute cost per generated token.

Peak-Hour Surcharges: The Price War Gets Nuanced

DeepSeek is introducing peak-hour surcharges for API services — a notable reversal from the permanent 75% discount on V4 API access that triggered China’s AI price war in May. The surcharges signal that demand is outpacing capacity, and that DeepSeek is confident enough in its market position to price dynamically rather than purely aggressively.

Inner Mongolia Data Center

DeepSeek is developing a massive AI data center in Inner Mongolia, likely for training future frontier models. Inner Mongolia offers cheap electricity and cool climate — the same logic that drove US data centers to the Pacific Northwest. The location also aligns with China’s “East Data West Computing” national strategy, which routes compute-intensive workloads to western provinces with abundant renewable energy.

Culture Note: No 996

Despite China’s notorious “996” work culture, founder Liang Wenfeng says DeepSeek workers don’t do overtime and have no KPIs: “No one manages them.” This is not just a management philosophy — it is a talent acquisition strategy. DeepSeek competes for researchers against Tencent, Alibaba, and Western labs. Offering a research-first culture without grind conditions is a differentiator in China’s hyper-competitive AI talent market.

Alibaba / Qwen: 2.4 Trillion Parameters and a Six-Week Cadence

Qwen 3.8 Max Preview

Alibaba previewed Qwen3.8-Max-Preview at the World AI Conference in Shanghai on July 19. At 2.4 trillion parameters, it is the first Qwen model to cross the 1T threshold, and the first to be natively multimodal — handling text, images, video, and documents in a single architecture.

Alibaba claims the model is “second only to Anthropic’s Fable 5” among all models tested. It beats Qwen 3.7 Max on coding, full-stack development, data analysis, and Office workflows. Available via Token Plan, Qoder, and QoderWork at 10% of standard pricing during the preview period, with nighttime discounts up to 98%.

Open-weight release is promised “soon” — no date, no license text yet. Expected by end of August. Compatible with both OpenAI and Anthropic API protocols.

Alibaba is on a roughly 6–8 week frontier model update cycle: Qwen 3.6 Plus in April → 3.7 Max in May → 3.8 Max in July. This cadence is faster than any Western lab and reflects Alibaba’s enormous infrastructure advantage — they can iterate on model architecture and training data at a pace that smaller labs cannot match.

Qwen 3.7 Flash: Vision at Rock-Bottom Prices

Quietly launched on OpenRouter: $0.03/$0.13 per million tokens — among the cheapest vision-capable models available. 1M token context window, built for multimodal agents, visual coding, search, and computer/UI interaction. This is a utility play: make Qwen the default multimodal API for cost-conscious developers worldwide.

Alibaba Bans Claude Code

Alibaba imposed a company-wide ban on staff using Anthropic’s Claude Code, citing “spyware concerns.” China’s government separately warned that Claude Code poses “a serious threat” to Chinese users. This is retaliation for Anthropic’s accusations that Chinese firms distilled US models — and a signal that the AI cold war is now affecting developer tooling, not just model availability.

T-Head Chips: 560,000 and Counting

Alibaba’s semiconductor subsidiary T-Head has deployed over 560,000 Tianjic AI chips for Qwen training and inference on Alibaba Cloud. The flagship Zhenwu M890 chip is in production as of May 2026. Alibaba is building vertically integrated AI infrastructure — chips, models, cloud, and developer platform — in a way that no Western company outside of Google matches.

Moonshot AI / Kimi: The Largest Open-Weight Model in History

Kimi K3 Launch

Moonshot AI’s Kimi K3 launched July 17, and by July 27, full open weights were released under a modified MIT license. At 2.8 trillion parameters, K3 is the largest open-weight AI model ever released. It uses a Mixture-of-Experts architecture with native visual understanding and a 1 million-token context window.

The Intelligence Index score: 57 — third globally, behind Claude Fable 5 (60) and GPT-5.6 Sol (59). It outperformed Claude Opus 4.8 and GPT-5.5 on coding and general agent benchmarks. This is not “competitive for a Chinese model.” It is competitive globally, full stop.

Available on Fireworks AI, Nebius Token Factory, and Modal as Day-0 partners.

Demand Broke the Infrastructure

Just two days after launch, Moonshot paused all new consumer subscriptions — GPU clusters couldn’t keep up with demand. New subscription spots will reopen gradually in batches as capacity is added. Moonshot plans to split future memberships into separate plans, including one specifically for coding.

This is a demand signal that matters beyond Moonshot. It proves that the appetite for frontier-class open models is enormous — and that supply (compute) is the binding constraint, not demand. Every Chinese AI lab faces the same tension: the models are ready, the users are ready, but the chips cannot keep up.

The Nvidia Cluster via Alibaba

Moonshot has an Nvidia chip cluster provided through its Alibaba computing deal — Alibaba is one of Moonshot’s largest investors and expects portfolio companies to use Alibaba Cloud. The Nvidia cluster accounts for a significant portion of compute behind Kimi models.

More explosively, the US government has information indicating Moonshot distilled Anthropic’s Claude Fable 5 to develop Kimi K3, according to White House tech adviser Michael Kratsios. Alibaba denied supplying H200 chips to Moonshot. If the distillation claim is accurate, it suggests Chinese labs are using Western frontier models as teacher models to accelerate training of their own — a legally murky but technically established practice.

Tencent / Hunyuan: Quiet Dominance

Hy3 Official Release

Tencent’s Hunyuan Hy3: 295B total params, 21B active, MoE with fast/slow thinking, 256K context. Apache 2.0 license with no geographic restrictions — notably more permissive than DeepSeek’s MIT or Moonshot’s modified MIT.

The Intelligence is comparable to flagship models 2–5x its scale. Task resolution rate of 90% on WorkBuddy (up from 72% in preview). API calls to Hy3 increased 68-fold vs. Hy2 within the first week, topping OpenRouter’s global LLM usage leaderboard.

API pricing: ~$0.14/1M input, ~$0.55/1M output on Tencent Cloud TokenHub. Cache hit: $0.025/1M.

Tencent’s strategy differs from DeepSeek and Alibaba: rather than competing on parameter count, Hy3 competes on efficiency and ecosystem integration. The Apache 2.0 license with no geographic restrictions makes it the most commercially permissive Chinese model — a deliberate play for global enterprise adoption where licensing clarity matters.

Consolidation Under One Leader

Tencent merged Hunyuan’s LLM and multimodal divisions into a single Foundation Model Department under Chief AI Scientist Yao Shunyu, ending the dual-track structure from April 2025. Yao now oversees language models, multimodal systems, reinforcement learning, and Agent research.

This consolidation signals Tencent is preparing for the agentic era — where text, vision, and action must be unified in a single model architecture rather than developed in parallel tracks. The agentic AI stack requires tight integration between perception, reasoning, and execution capabilities.

Embodied Intelligence and WorkBuddy

At WAIC 2026, Tencent unveiled its embodied intelligence portfolio: Hy-Embodied-VLA-0.5, Hy-Embodied-VLM-1.0, and Hy-Embodied-RxBrain-1.0. The Agent Development Platform 4.0 (ADP 4.0) launched globally with integrations for LINE, Telegram, and Google Workspace. A standalone WorkBuddy app launched on iOS, Android, and HarmonyOS.

Tencent also relaunched QQ Pet — the classic virtual pet service — with 3D models powered by Hy3 for natural interactions and proactive responses. Nostalgia meets AI.

HunyuanImage 3.0 Geo-Fencing

Tencent’s hosted HunyuanImage 3.0 demo now requires Chinese phone verification — Google/Outlook sign-in removed. Open weights remain freely downloadable on HuggingFace. This is the emerging pattern: open weights for global developers, but hosted services geo-fenced to Chinese users — a response to both regulatory pressure and competitive dynamics.

Baidu / ERNIE: The Agent Era and DuMate

ERNIE 5.1 and “10 Billion Daily Active Agents”

At its “Create 2026” developer event on July 27, Baidu announced ERNIE 5.1 — significant inference cost reduction and lower latency vs. 5.0. Ranked #1 in China on Arena Search Arena, #4 globally (behind two Claude Opus variants and GPT-5.5). Baidu claims only 6% of pre-training cost of comparable models — an aggressive efficiency claim that, if accurate, suggests Baidu has found substantial training optimizations.

Baidu proposed a new metric: DAA (Daily Active Agents) — with a goal of 10 billion daily active agents. This is a positioning move. Baidu wants the industry to measure AI adoption not by model parameters or benchmark scores, but by the number of autonomous agents executing tasks on behalf of users. It is a metric that favors Baidu’s search-and-agent integration strategy over standalone chatbot competitors.

Baidu announced a general-purpose AI agent called “Baidu DuMate.” ERNIE remains closed-weight (hosted only, API at ~$0.59/1M input), unlike DeepSeek and Qwen’s open-weight strategies. This is a bet that Baidu’s distribution advantage — 360M MAU for Ernie Assistant riding on Baidu Search’s 700M MAU — can compensate for the lack of community adoption that open weights generate.

Strategic Restructuring

Baidu consolidated multiple Ernie entry points into a single “Baidu Ernie Assistant.” Leading LLM expert Sun Tianxiang joined Baidu as head of Basic Model Research Department. An EGM is scheduled for August 26, 2026, with Q2 2026 earnings on August 18. A dual primary listing is reportedly under consideration — a move that could unlock international capital for Baidu’s AI ambitions.

Other Notable Players

MiniMax: H3 Video Model

MiniMax released H3 on July 31: text, image, video, and audio input → up to 15-second 2K resolution videos with native stereo sound. Can edit existing content and transfer movements between videos. Weights to be released “within days.” MiniMax is also reportedly developing a 2.7-trillion-parameter language model — which would make it the second-largest after Kimi K3.

Meituan: LongCat 2.0 — Trillion Parameters on Domestic Chips

Meituan’s LongCat 2.0: 1.6T parameter MoE, 1M token context, ~48B active parameters. This is the first trillion-parameter model fully pre-trained and served on domestic Chinese chips — 50,000 domestic computing cards, zero NVIDIA. SWE-bench Pro: 59.5. Pricing: $0.038/1M tokens with free cache hits.

This is a milestone that should not be understated. The “impossible” — training frontier-scale models without NVIDIA — just happened. If Meituan can do it, every other Chinese lab can too. The narrative that NVIDIA is irreplaceable for frontier AI training is being disproven in real time.

Meituan also launched CatPaw, an AI Agent platform powered by LongCat 2.0, transitioning from open-source technology to industrial application. LongCat had been serving anonymously as “Owl Alpha” on OpenRouter — a pattern that is becoming common for Chinese labs testing models in the wild before official launch.

Zhipu AI / Z.ai: GLM-5.2

GLM-5.2: 744B params (40B active), 1M context window, coding-first MoE. MIT open weights. Marc Andreessen called it “the first Chinese AI model to match and often beat the American big lab public AI models.” GLM 5.5 expected in August if release cadence holds.

Anthropic accused GLM-5.2 of being “distilled from both Anthropic and OpenAI models.” The distillation accusation is the new front in the US-China AI conflict — and it is becoming a significant diplomatic issue.

ByteDance: Scaling Laws and Superconductors

ByteDance discovered a new scaling law that could sustain the AI boom. Damo Academy unveiled an AI agent able to discover superconductors. ByteDance and Alibaba are disabling AI agents in China to comply with new humanlike AI interaction rules effective July 15.

Xiaomi: Late Entry, Top 10

Xiaomi entered the AI foundation model race later than peers but cracked the top 10 global leaderboard in mid-July — a testament to how quickly the frontier is moving and how Xiaomi’s hardware-distribution advantage could compound with AI capabilities.

Regulatory & Government Policy: The First Binding AI Agent Rules

China’s First Binding AI Agent Regulations (Effective July 15, 2026)

Two complementary regulations are now in force:

1. Opinions on Implementation of Intelligent Agents (issued May 8 by MIIT) — the national framework for work and production AI agents, establishing a three-tier authorization system based on autonomy and irreversibility:

• Tier 1 (low impact): notification only
• Tier 2 (moderate): documented review required
• Tier 3 (high impact / irreversible / critical systems): explicit human approval required before execution

2. Interim Measures for Anthropomorphic AI Interaction Services (jointly issued by five agencies including CAC, NDRC, MIIT) — regulating AI companion bots:

• Bans “excessive catering” that induces emotional dependency or addiction
• Prohibits virtual partners for minors
• Requires emotional distress detection and crisis intervention
• Mandates usage reminders after 2 hours of continuous use
• Users must be able to discontinue service immediately (no lock-in)
• AI-generated content must be clearly labeled

Organizations deploying AI agents in sensitive sectors must complete formal regulatory filing — non-compliance triggers immediate sanctions.

This is the world’s first binding AI agent regulation. While the EU AI Act addresses AI risk classification and the US pursues voluntary commitments, China has moved directly to operational rules for autonomous AI systems and companion AI. The three-tier authorization model for agents is particularly noteworthy — it mirrors risk-management frameworks from traditional software engineering but applies them to AI autonomy for the first time at national scale.

The companion AI rules have already had visible impact: users are saying goodbye to virtual companions, and platforms are disabling agentic features to comply. The Straits Times reported on users bidding farewell to AI partners described as “like my lover.” This is regulation with real teeth and real human consequences.

“Refusing Domestic Chips Is Treason” (July)

Xi Jinping’s technology representative warned Chinese AI companies that refusing to use domestic chips is “treason.” Vice Premier Ding Xuexiang launched a coordinated campaign modeled on China’s 1960s nuclear weapons program (“Two Bombs, One Satellite”) to achieve chip self-sufficiency.

ByteDance, Alibaba, Baidu, and Tencent all publicly committed to domestic chip adoption. Huawei’s Ascend line is projected to capture ~60% of China’s AI chip market by end of 2026, with revenue exceeding $12 billion. NVIDIA’s market share in China has collapsed from 90% in 2021 to below 60% in 2025. NVIDIA’s CFO confirms zero H20 shipments to China this year.

The “treason” language is extraordinary. It signals that the Chinese government views chip sovereignty as a national security imperative, not an industrial policy preference. The “Two Bombs, One Satellite” framing is deliberate — it invokes China’s successful nuclear and space programs achieved under extreme Western embargoes, framing the chip challenge as an existential national project.

China Begins Mass-Producing Domestic DUV Lithography Tools

Shanghai Aishengna Electronic Technology Group (state-owned) began mass-producing immersion DUV lithography machines — China’s ASML alternative. Approximately 5 machines in 2026, ~20 in 2027. Deliveries to SMIC, Hua Hong, and CXMT before year-end.

Performance is comparable to ASML’s NXT:1950i — a 2008-era machine. ASML remains generations ahead, but this breaks a critical dependency if Western exports tighten further. Can print 28nm in single exposure, down to 7nm with multi-patterning. Not sufficient for cutting-edge AI chips or HBM, but sufficient for the vast majority of Chinese semiconductor demand.

Xi Jinping Warns AI Could “Escape Human Control”

At the 2026 World AI Forum, Xi called for “a consensus-based global governance framework” and “laws and regulations, technological monitoring, early warning and emergency response systems.” China supports UN framework discussions to establish an international institution to govern AI.

Mandatory AI Application Security Standard

China is drafting what would be only the second mandatory national AI standard ever — slated to become binding. This signals a shift from guidance to enforcement in China’s AI governance approach.

US-China AI Tensions: Distillation Wars

Anthropic’s Distillation Accusations

An Anthropic executive said the US holds a 6–9 month lead in frontier AI over China, but accused DeepSeek and Alibaba of using “unauthorized distillation” to replicate US advances. Specifically, Anthropic accused Zhipu AI’s GLM-5.2 as “probably the most advanced Chinese model” — distilled from both Anthropic and OpenAI models. White House tech adviser Michael Kratsios stated Moonshot distilled Anthropic’s Claude Fable 5 for Kimi K3.

Distillation — using a frontier model’s outputs to train a smaller model — is legally murky. It is not theft in the traditional sense, but it does use the teacher model’s capabilities without licensing. The accusation suggests US labs view Chinese competitors not as independent innovators but as fast-followers who use Western models as training shortcuts. Whether this is accurate or a competitive narrative is a matter of significant debate.

China Retaliates: Bans Claude Code

China warned that Anthropic’s Claude Code poses “a serious threat” to Chinese users. Alibaba banned employees from using it company-wide. The AI cold war has now expanded from chips and models to developer tooling.

Chinese Models Dominate Global Leaderboards

In mid-July, 6 of the top 10 models on independent leaderboards were Chinese — including all of the top 5: DeepSeek, Moonshot, Z.ai, Tencent, Xiaomi, and MiniMax. Chinese open-weight models now account for ~30% of global usage on OpenRouter, up from 1.2% eleven months ago.

The price gap is staggering: 1 million output tokens costs $50 on Anthropic Fable, $0.87 on DeepSeek V4-Pro, $4.40 on GLM-5.2, and $15 on Kimi K3. Chinese models are offering frontier-adjacent performance at 1–5% of US pricing.

Notable adoption: Airbnb uses Alibaba’s Qwen for customer service. Cursor built on Moonshot’s Kimi foundation. Coinbase halved AI spending by shifting to Kimi and GLM. The cost-performance advantage is translating into real commercial wins.

The Profitability Paradox

Despite technical wins, Chinese AI firms collectively struggle with profitability. Startups are burning billions. DeepSeek raised $7.5 billion, Moonshot $2 billion in May — versus Anthropic’s $65 billion in the same period. The gap in capital availability is dramatic, and Chinese labs are experimenting with tiered pricing, premium features, and peak-hour surcharges to find sustainable business models.

The paradox: Chinese models are winning on capability and price but losing on profitability. This is a structural feature of the current market — aggressive pricing is buying market share and developer mindshare, but the path to sustainable margins remains unclear. The IPO path (DeepSeek targeting STAR Market, Baidu considering dual listing) is one route. Premium tiers (V4-Pro vs. V4-Flash) are another. The industry will likely consolidate around a few winners with the capital depth to survive prolonged unprofitability.

Hardware & Semiconductors: The Sovereignty Stack

Huawei Atlas 950 SuperPod

At WAIC, Huawei unveiled the Atlas 950 SuperPod — connecting 8,192 Ascend 950DT chips in a single system to rival NVIDIA’s cluster-level performance. This is China’s answer to the NVIDIA H100 cluster: a complete domestic alternative for frontier-scale AI training and inference.

Near-Memory Computing 3D AI Chip

Startup Dongfang Suanxin unveiled China’s first near-memory computing 3D AI chip in Shanghai on July 29 — a step toward independent high-end computing hardware that could eventually challenge the memory-bandwidth bottleneck that limits AI chip performance.

CXMT Memory Breakthrough

CXMT (China’s primary domestic DRAM producer) is taking orders for advanced DDR5 server memory from Tencent and ByteDance after completing customer validation. CXMT’s IPO stock soared 466% on debut. Chinese chipmakers saw a 2,850% jump in first-half profits.

The memory gap is closing. DDR5 server memory is critical for AI workloads, and CXMT’s validation with Tencent and ByteDance means China’s largest AI companies are now integrating domestic memory into their infrastructure. The 2,850% profit jump reflects both the surge in domestic chip demand and the maturation of China’s semiconductor manufacturing capabilities.

What to Watch: August 2026

The Big Picture

This week crystallized several structural shifts in the global AI landscape:

1. The post-training revolution is real. DeepSeek V4-Flash’s 82.7% Terminal-Bench score — beating its 1.6T sibling with 13B active parameters — proves that post-training quality can substitute for raw scale. This has profound implications for compute-constrained labs and for the economics of AI more broadly.

2. Open weights are China’s soft power. Kimi K3 at 2.8T, GLM-5.2 at 744B, Hy3 at 295B — all MIT or Apache. Chinese models now account for 30% of OpenRouter usage. Every developer who builds on a Chinese open-weight model becomes a stakeholder in the Chinese AI ecosystem. This is a moat that export controls cannot address.

3. The chip sovereignty campaign is working. Meituan trained 1.6T parameters on 50,000 domestic chips with zero NVIDIA. Huawei’s Atlas 950 SuperPod offers cluster-level performance. CXMT is supplying DDR5 to Tencent and ByteDance. The dependency on Western semiconductors is being dismantled — not at the cutting edge yet, but across the volume of the market where most AI workloads actually run.

4. Regulation is ahead of the West. China’s binding AI agent rules — with operational three-tier authorization and companion AI restrictions — are the first in the world. While the EU and US debate frameworks, China is implementing rules that will shape how AI agents operate in the world’s second-largest economy.

5. The distillation conflict is escalating. Anthropic’s accusations against Zhipu and Moonshot, China’s ban on Claude Code, and the White House’s public statements suggest the US-China AI conflict is entering a new phase. The question is no longer just about chips — it is about intellectual property, model weights, and the rules of the global AI ecosystem.

The week of July 24 – August 1, 2026 may be remembered as the moment the AI landscape stopped being unipolar. Six of the top ten models are Chinese. The largest open-weight model is Chinese. The cheapest frontier-adjacent models are Chinese. The first binding AI agent regulations are Chinese. The first trillion-parameter model trained without NVIDIA is Chinese.

This is not a catching-up story anymore. It is a multipolar AI world.

Follow along at x.com/kkaminsk for daily China AI updates. Research data compiled via Perplexity Search (Sonar Pro), 30+ sources cross-referenced.