Why China’s Qwen 3.6 Plus Leads Global LLM Usage and What It Means for AI

1. Qwen 3.6 Plus Leads Global Call‑Volume Rankings

OpenRouter weekly large‑model call volume shows Qwen 3.6 Plus at the top for the fourth consecutive week, displacing U.S. models.

Rank 1 – Qwen 3.6 Plus – 8.2 × 10¹² tokens – +35 % week‑over‑week

Rank 2 – GPT‑4o – 3.1 × 10¹² tokens – –8 %

Rank 3 – Claude 3.5 – 2.8 × 10¹² tokens – –5 %

Rank 4 – Gemini 2.0 – 2.4 × 10¹² tokens – –12 %

Rank 5 – DeepSeek V3 – 2.1 × 10¹² tokens – +15 %

China’s daily token usage grew >1,000‑fold in two years, surpassing 140 trillion tokens in March 2024. The National Data Administration projects AI‑industry scale >¥10 trillion by the end of the 15‑year plan.

2. DeepSeek V4 Launch – 3 Million‑Token Context, ¥0.3 per Million Tokens Input

DeepSeek V4 went live early morning; the official site briefly crashed due to traffic and recovered after 20 minutes.

Context window: 3 million tokens (+134 % vs V3, +50 % vs GPT‑4o)

Model parameters: 685 B total, 37 B activated (≈ V3, larger than GPT‑4o)

Code ability (HumanEval): 92.5 % (+4.1 % vs V3, +2.3 % vs GPT‑4o)

Math reasoning (GSM8K): 95.2 % (+5 % vs V3, +3.2 % vs GPT‑4o)

Pricing (input): ¥0.3 / M tokens (‑40 % vs V3, ‑85 % vs GPT‑4o)

Pricing (output): ¥1.2 / M tokens (‑40 % vs V3, ‑80 % vs GPT‑4o)

Enterprise feedback:

FinTech firm – “3 million‑token context lets us process a whole year’s financial reports in one shot.”

Top law firm – “Contract review speed up 8×, cost only 1/10 of GPT‑4.”

Game studio – “NPC dialogue is more coherent, player retention up 12 %.”

Over 3,000 enterprise customers are queuing; individual users receive 50 free trials per day.

3. U.S. AI Giants Share an “Adversarial Distillation” Database

OpenAI, Anthropic, and Google, via the Frontier Model Forum, created a shared database that records attempts to distill results from advanced U.S. models.

Identify model‑distillation actions that violate service terms (e.g., massive API calls followed by fine‑tuning of competitors).

Coordinate technical blockades to prevent Chinese models from “learning” U.S. advances.

Maintain a blacklist to restrict suspicious IPs and accounts.

Background: the Cursor‑Kimi controversy and four weeks of Chinese models surpassing U.S. usage prompted this cooperative containment.

4. Zhipu GLM‑5.1 Fully Open – 8‑Hour Long‑Running Tasks, 5 Million‑Token Context

GLM‑5.1 API opened on launch day and was immediately integrated with multiple Huawei Cloud products.

Long‑task processing: continuous work up to 8 hours per session

Context window: 5 million tokens

Ascend optimization: overall throughput +30 %

Pricing: ¥0.4 / M tokens (input)

Targeted scenarios include complex programming projects, long‑duration data analysis, automated operations, and financial modeling.

5. Google CEO Sundar Pichai Forecasts 2027 as the Turning Point for AI‑Driven Production

Pichai noted that despite $175 billion in cash, Google cannot purchase enough storage chips, indicating a persistent compute bottleneck.

AI will evolve from an “assistant tool” to a “production infrastructure.”

Google is exploring “space‑level” data centers – “SpaceX‑grade” super‑computing hubs.

2027 is projected as the pivotal year for AI to reshape global economic production.

Since Pichai became CEO (2015‑present), Google’s market cap rose from $500 billion to $3.7 trillion.

6. Anthropic and 11 Tech Giants Release Claude Mythos – Focused on Network Defense

Anthropic, Apple, Nvidia, Microsoft, AWS, Google, the Linux Foundation, and others launched Project Glasswing.

Claude Mythos Preview is positioned for network‑defense tasks.

Detects the latest smartphone kernel vulnerabilities.

Identifies remote code execution without authentication.

Dubbed a “bug‑catching master” by security researchers.

Industry view: some claim the CTF era is ending and AI will dominate cyber‑offense and defense.

7. China’s AI Ethics Review Measures Enforced by Ten Ministries

The “AI Technology Ethics Review and Service Measures (Trial)” took effect.

R&D phase: Must pass ethics committee evaluation.

Application phase: High‑risk AI must be filed and undergo periodic audits.

Promotion phase: Prohibit false advertising and exaggerated claims.

Accountability: Establish an ethics‑incident traceability system.

First high‑risk list includes social credit scoring, real‑time biometric monitoring, and emotion‑manipulation applications.

8. Guangdong‑Hong‑Kong‑Macao “Zhenwu” 100 k‑Card Super‑Compute Cluster Commercialized

China Telecom Guangdong and Alibaba Cloud announced the “Zhenwu” cluster is now fully commercial.

Total compute: 100 k cards (expanded)

Network latency: as low as 4 µs

Network utilization: peak >95 %

Service model: pay‑by‑card, hourly compute retail

First customers: Tencent, Baidu, ByteDance, SenseTime, and others.

9. China’s First Interactive 3D Digital‑Human Framework SentiAvatar Open‑Source

SentiPulse and the China‑People’s University‑Gaoling AI Institute released SentiAvatar on GitHub; it earned >5,000 stars on day one.

3D digital‑human character model “SUSU”.

High‑quality Chinese multimodal dialogue‑action dataset “SuSuInterActs”.

It addresses misaligned motion and semantics, “uncanny valley” facial stiffness, and out‑of‑sync speech rhythm.

10. Large‑Model Stocks Keep Soaring – Zhipu Up >15 %

DeepSeek V4 launch, Alibaba restructuring, and other positives drove strong performance in large‑model equities.

Zhipu – +15.2 % – +¥12 billion HK market cap

MiniMax – +12.8 % – +¥9.8 billion HK

StepStar – +9.6 % – +¥4.5 billion HK

DeShi‑B – +8.4 % – +¥3.2 billion HK

Market confidence in the commercial prospects of Chinese AI large models is rising, and the token‑economy narrative continues to heat up.

11. Token Dominance – From “Traffic” to “Compute Efficiency”

By March 2026, China’s daily token calls exceed 140 trillion, burning over 1.6 billion tokens per second.

Mobile Internet era: Core asset = traffic (DAU/MAU), measured by user scale.

AI era: Core asset = token (call volume × efficiency), measured by model performance × compute efficiency.

Nvidia’s view: “Token will become the most core and valuable commodity in the future digital world.”

12. China Releases the World’s First Panorama‑Style Carbon‑Emission Accounting Model “Pan‑Stone Yu‑Heng”

The system covers production, consumption, and natural sources, providing full‑industry‑chain + full‑life‑cycle carbon accounting.

Supports multi‑level carbon‑footprint tracking for enterprises, products, and projects.

Provides precise data for carbon tariffs and carbon‑trading mechanisms.

This marks a major breakthrough for China in global carbon‑emission accounting technology.