📋 Table of Contents
Jump to any section (15 sections available)
📹 Watch the Complete Video Tutorial
📺 Title: The Chinese AI Iceberg
⏱️ Duration: 1626
👤 Channel: bycloud
🎯 Topic: Chinese Iceberg
💡 This comprehensive article is based on the tutorial above. Watch the video for visual demonstrations and detailed explanations.
In 2025, the global AI landscape is undergoing a seismic shift—Chinese AI labs now dominate open-source innovation, rivaling and in some cases surpassing private models from the U.S. From stealth startups to tech giants, China’s AI “iceberg” runs deep, with layers of groundbreaking research, model releases, and infrastructure development happening beneath the surface. This comprehensive guide dives into every level of that iceberg, revealing the companies, models, breakthroughs, and even controversies shaping the future of artificial intelligence.
Whether you’re a developer, entrepreneur, researcher, or tech strategist, understanding this ecosystem isn’t optional—it’s essential for navigating the next wave of AI disruption. We’ll explore everything from DeepSeek’s underdog revolution to Huawei’s scandal, and from Alibaba’s Quinn dominance to underground labs pushing trillion-parameter models.
The Iceberg Framework: Mapping China’s AI Ecosystem
The speaker uses an “iceberg” metaphor to categorize Chinese AI players by visibility and influence:
- Surface (Mainstream): Widely known commercial labs
- Mid-level (Big Three): Alibaba, ByteDance, Tencent
- Startups & Rising Stars: Moonshot, Zhipu AI, MiniMax
- Pre-DeepSeek Era: Early movers like Baidu and 01.ai
- Underground Labs: Niche but high-impact researchers
- Deepest Layer: Unexpected entrants (e.g., social platforms)
Level 1: The Blue Whale – DeepSeek’s Open-Source Revolution
DeepSeek stands as the most unconventional force in open-source AI. Despite being a self-funded startup with minimal compute compared to giants, it has repeatedly disrupted the market.
Key Achievements & Philosophy
- Releases all infrastructure and training scripts openly—a scale unmatched by any other lab
- Never self-hypes: Social media posts only say, “We made a significant performance boost.”
- Won Best Paper Award at ACL for Native Sparse Attention
- Caused Nvidia stock to drop 10% upon release of DeepSeek R1 due to market fears
- DeepSeek V3 3.1 now natively supports Chinese hardware—potentially a game-changer for domestic AI infrastructure
Level 2: The Chinese Big Three – Tech Giants with AI Ambitions
Alibaba Cloud & the Qwen (Quinn) Dynasty
Alibaba Cloud—China’s answer to Amazon Web Services—has cemented Qwen as the backbone of open-source LLM research.
- Qwen 3 235B 2507 Reasoning is currently the #1 open-source model on benchmarks
- Offers a full spectrum of model sizes: 7B for mobile, mid-sized for research, large for production
- Released 100+ open-weight checkpoints since 2023 with 100M+ total downloads
- Also leads in video generation: Qwen-VL (closed-source) ranks 7th globally; its open-source variant is the best available
- Adopted by top universities in both China and the U.S. for cutting-edge research
ByteDance & ByteDance AI (ByteSeed)
Though late to the AI race, ByteDance exploded onto the scene in 2025 with world-class generative models.
- Seed Thinking v1.5: Briefly held the title of best Chinese reasoning model (though closed-source)
- Seed Dream 3.0: #1 on Image Arena, beating Google, Black Forest Labs, and Flux
- Seed Video 1.0: #1 on Video Arena—a stunning leap from nowhere
- Plans to invest $20 billion in AI compute, including Chinese RAM infrastructure
- Behind Dobao, China’s #2 AI assistant app (as of March 2025), featuring:
- Text-to-speech
- Image & video generation
- Multimodal understanding
- Voice cloning
Tencent & HunYuan
Tencent entered AI cautiously but is now integrating models deeply into its ecosystem.
- Launched HunYuan Large in late 2024—the largest open-source model at the time
- Flagship reasoning model: HunYuan T1 (minimal technical disclosure)
- Released HunYuan Turbo S, a Mamba-Transformer hybrid, with a detailed research paper
- Active in 3D and video generation; latest: text-to-360° world model
- Fully integrated into WeChat, giving it massive user reach
Level 3: Rising AI Startups – The New Vanguard
Moonshot AI
Founded in April 2023 by Yangzhi Yang—lead author of Transformer-XL (5,000+ citations) and XLNet (11,000+ citations), ex-Meta AI and Google Brain.
- Raised $200M at a $300M valuation within 2 months of founding
- Released Kimi K2, which briefly held the non-reasoning open-weights crown
- First AI app to support all 200,000 Chinese characters
- Office culture: Meeting rooms named after iconic artists, album covers on walls
Zhipu AI (ZAI)
One of China’s earliest AI startups, evolving from Tsinghua University researchers in 2019.
- GLM4.5 hybrid reasoning model: Top 3 on open-source LLM leaderboard
- Outperformed Kimi K2 with 30% fewer parameters
- Created ChatGLM (June 2024), an open-source chat model with 41,000+ GitHub stars
- Built the first transformer-based text-to-video generator
MiniMax
A stealth powerhouse with strengths in multimodal AI and speech.
- Launched Glow (AI roleplay chatbot) in 2021—5M+ downloads
- Developed Hyo AI, a multimodal chat app
- Released three major open models in 2025:
- MiniMax Text: Linear attention LM
- MiniMax VLM: Vision-language model
- MiniMax M1: Hybrid attention model with 1M context window (#7 on leaderboard)
- Best-in-class text-to-speech: #1 and #3 on Artificial Analysis TTS Arena
- Just released MiniMax M2 (June 2025)—tops open-weights leaderboard, ranked #9 globally on Artificial Analysis Intelligence Index
Qwen (Not Alibaba’s Qwen!) – The Video Specialist
Often confused with Alibaba’s model, this Qwen** (from company **Qwen.ai)** focuses solely on video.
- Released Cling video model in mid-2024
- Still ranks among the top 3 video generators, behind only ByteSeed and Google
- No language model or research publications—pure video focus
- Parent company operates in TikTok-like short-video space
Honorable Mention: Manis AI – The Agentic Phenomenon
Though now Singapore-based (closed Chinese offices in July 2025), Manis AI exploded in March 2025.
- Launched an autonomous agent app with a virtual computer environment
- Can research, calculate, and surf the web independently
- Accumulated 2M+ waitlist users before public launch in May 2025
- Faced backlash over influencer-only early access, but claimed it was due to server scaling limits
- Now faces existential threat as big labs enter the agentic application space
Level 4: The Pre-DeepSeek Era – Early Pioneers (2023–2024)
Baidu & ERNIE
- Launched ERNIE 3.0 in July 2021—based on GPT-3 architecture (175B params)
- Remained closed-source until July 2025, when it released ERNIE 4.5
- Includes a rare Mixture-of-Experts Vision LM
- Not yet on major leaderboards, but worth testing
01.ai – The Ghost Lab
- Founded by Kai-Fu Lee (ex-Google China head)
- Dropped a near-SOTA open model in November 2023 with zero prior track record
- Briefly closed the gap with GPT-4, even winning on key benchmarks
- Released private model Yi-Large in May 2024
- Went completely silent after January 2025—no updates, no news
BAIIN (BRUN)
- Released Byron 3 (Jan 2024) and Byron 4 (May 2024)
- Byron 4 briefly led Chinese AI rankings before Alibaba’s Qwen 2 dethroned it
- Also vanished from public view—possibly overwhelmed by Qwen’s dominance
Level 5: The Underground Powerhouses – Hidden Gems
StepFun
- First Chinese startup to build a 1T-parameter LLM (Step 2, Nov 2024)
- Step 3: Open-source multimodal reasoning model
- Vision understanding is near SOTA, but only 4,400 downloads due to size and complexity
OpenBMB (Open Lab for Big Model Base)
- Unique approach: Publishes datasets AND fine-tuned models—rare in the field
- MiniCPM and MiniCPM-V series: Optimized for performance and efficient deployment
- MiniCPM-V 2.6 rivals GPT-4o in language + vision, fully open-source
- Pioneered novel techniques like RLPR (Reinforcement Learning from Process Rewards)
Huawei & the Pangu Scandal
A major controversy erupted in 2025 involving Huawei’s AI lab.
- Whistleblower revealed that Pangu models were rebranded open-source models:
- Pangu = fine-tuned Qwen 1.5 10B
- Pangu = fine-tuned Qwen 2.5 14B
- Pangu = fine-tuned DeepSeek V3
- Models were padded with dummy parameters to match advertised sizes
- Internal team failed to train original models on Ascend chips (38B, 135B, 718B attempts)
- A politically favored internal lab stole data, faked reports, and claimed bonuses
- GitHub exposé received 11,000+ stars—reputational damage likely severe
SenseTime & SenseNova
- Founded by creators of DeepID (2014 facial recognition breakthrough)
- Now works on autonomous vehicles, medical AI, and sensors
- SenseNova V6.5 (2025): Multimodal model claiming to beat Gemini 2.5 Pro and Claude 3.5 Sonnet
- But: No public access—claims unverifiable
Shanghai AI Laboratory
- Government-backed R&D hub (founded July 2020)
- Connects top talent from China’s elite universities
- Produces ~20 top-tier AI papers per month
- InternLM series: Collection of small open models
- Intern S1 (July 2025): Built on Qwen 3 235B + InternViT vision encoder
- Matches DeepSeek R1 in language, adds unique vision capabilities
Ant Group (Alibaba’s Fintech Arm)
- Behind Alipay; now deeply invested in AI research
- Pioneered Large Language Diffusion Models and early fusion techniques
- September 2025: Released Ring 1T—a 1-trillion-parameter reasoning model
- Launched Inclusion AI initiative for all open-source releases
- Effectively gives Alibaba two SOTA AI labs (Cloud + Ant Group)
Essential Tool: Cherry Studio – The Universal AI Gateway
Not a model lab, but a critical infrastructure piece.
- Open-source app (32,000+ GitHub stars) by Ching Huay Technology
- Connects all major LLM providers—Chinese and global
- Works on all operating systems
- Features clean UI, native tool integration, and easy model switching
- Bridges the gap between walled U.S. ecosystems and China’s AI landscape
Level 6: The Deepest Layer – Unexpected AI Entrants
Even non-tech companies are jumping in.
Xiaohongshu (Red Note)
- Chinese social platform (alternative to TikTok during U.S. ban)
- No prior AI history
- June 2025: Launched DOTS LM1 under “Retino High Lab”
- 142B total parameters, 14B active (MoE architecture)
- Shows promise on cost-performance metrics
Why Most AI Implementations Fail – Critical Insights
Referencing HubSpot’s research compilation (featuring OpenAI, Anthropic, Atlassian):
- AI fails when treated as a standalone tool rather than an agentic system
- Consistent memory is essential for real-world task completion
- Evaluation must be business-relevant, not just academic
- Successful deployments (e.g., HubSpot) track AI agents like human employees—measuring resolution rates, speed, and ROI
Timeline of Key Chinese AI Milestones (2023–2025)
| Date | Event | Significance |
|---|---|---|
| Apr 2023 | Moonshot AI founded | Ex-Google/Meta AI lead enters race |
| Sep 2023 | Tencent launches HunYuan | First major LM from gaming giant |
| Nov 2023 | 01.ai releases Yi model | Closes gap with GPT-4 overnight |
| Jan 2024 | BAIIN launches Byron 3 | Early private model contender |
| Jun 2024 | ChatGLM released | 41K+ GitHub stars; bilingual chat focus |
| Jul 2024 | DeepSeek R1 released | Causes 10% Nvidia stock drop |
| Apr 2025 | ByteSeed dominates image/video arenas | #1 in both categories |
| Jul 2025 | Alibaba Qwen 3 235B released | Becomes #1 open-source model |
| Jun 2025 | MiniMax M2 tops leaderboard | New open-weights champion |
| Jul 2025 | Baidu releases ERNIE 4.5 | First open-source ERNIE after 4 years |
How to Track Chinese AI Developments
Stay ahead with these methods:
- Use Papers.ai Scout to filter arXiv uploads by institution (e.g., Shanghai AI Lab)
- Monitor Hugging Face leaderboards for open-weight models
- Follow Artificial Analysis for TTS, video, and multimodal rankings
- Install Cherry Studio to test models across providers
The Strategic Implications for Global AI
China’s approach differs fundamentally from the U.S.:
- Openness: Chinese labs lead in open-weight releases (Qwen, DeepSeek, GLM)
- Hardware Independence: DeepSeek V3’s native Chinese chip support signals decoupling from Nvidia
- Application Focus: From WeChat to Dobao, AI is embedded in daily digital life
- Compute Investment: ByteDance’s $20B AI spend shows long-term commitment
Conclusion: Navigating the Chinese AI Iceberg
The “Chinese Iceberg” is no longer hidden—it’s rising fast, with open-source innovation at its core. From DeepSeek’s scrappy transparency to Alibaba’s Qwen ecosystem, and from MiniMax’s multimodal mastery to the tragic implosion of Huawei’s Pangu, this landscape offers both opportunity and caution.
For businesses and developers, the key takeaways are clear:
- Embrace open models—they’re now competitive with or superior to closed alternatives
- Prioritize agentic capabilities over raw parameter counts
- Monitor underground labs—breakthroughs often emerge from unexpected places
- Beware of hype—verify claims (e.g., SenseTime) and watch for scandals (e.g., Huawei)
As the U.S. contemplates regulatory moves, China’s AI ecosystem continues accelerating—driven by a mix of state support, private investment, and a culture of open collaboration. The future of AI won’t be decided in Silicon Valley alone. It’s being coded, trained, and deployed across the entire Chinese Iceberg.
