📋 Table of Contents
Jump to any section (14 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.
As open-source AI models increasingly rival—and in some cases surpass—private systems, a quiet revolution is unfolding in China. From well-funded tech titans to stealthy startups and even government-backed research labs, Chinese AI development is accelerating at a pace that’s reshaping global benchmarks, stock markets, and the very definition of open innovation. This comprehensive guide dives deep into the “Chinese Iceberg”—a layered ecosystem of AI players you need to understand before the geopolitical and technological tides shift irreversibly.
Based entirely on an in-depth video transcript, this article extracts every insight, model name, company detail, performance metric, and strategic observation to deliver the most complete map yet of China’s AI landscape. Whether you’re an entrepreneur, developer, investor, or policy analyst, knowing who’s driving progress—and how—will be critical for your business, career, and understanding of the AI future.
The Iceberg Framework: Mapping China’s AI Layers
The “Chinese Iceberg” metaphor organizes AI players by visibility and influence:
- Surface Level: Mainstream, publicly recognized giants
- Middle Layers: Well-funded but less visible startups and research divisions
- Deep Layers: Underground labs, academic institutes, and stealth innovators
- Deepest Tier: Unexpected entrants with sudden AI pivots
Each layer contributes uniquely to China’s rapid AI ascent—often with open-source ethos, hardware sovereignty, and multimodal ambition.
HubSpot’s AI Predictions: Context for the Iceberg
Before diving in, the transcript highlights a critical external resource: HubSpot for Startups’ free report featuring unfiltered predictions from executives at OpenAI, Anthropic, Atlassian, and more. Key takeaways include:
- AI is evolving from a standalone tool into an autonomous personal assistant that completes full tasks
- Companies like HubSpot now track AI agents alongside human reps in sales dashboards, measuring resolution rates and speed
- Emerging markets adopt AI faster due to fewer legacy system constraints
- Success requires business-aligned evaluation—not just leaderboard scores
This context sets the stage: China’s AI iceberg isn’t just about models—it’s about deployable, measurable, and scalable intelligence.
Level 1: The Blue Whale – DeepSeek
At the surface sits DeepSeek, the “nonchalant top dog” of open-source AI and a true anomaly in the industry.
Unconventional Powerhouse
Despite being an unfunded startup run solely on CEO Highfire’s capital—and with far less compute than major labs—DeepSeek releases nearly everything open-source:
- Infrastructure code
- Training scripts
- State-of-the-art models
Their social media is famously understated, posting only lines like: “We made a significant performance boost.”
Market Impact & Technical Prowess
DeepSeek’s influence is undeniable:
- The release of DeepSeek R1 caused Nvidia’s stock to drop 10% due to Wall Street fears
- Their paper “Native Sparse Attention” won Best Paper at the Association for Computational Linguistics (ACL) conference
- DeepSeek V3 3.1 now natively supports Chinese AI hardware—a potential inflection point
Level 2: The Chinese Big Three
These tech giants rival OpenAI, Anthropic, and xAI in budget and ambition.
Alibaba Cloud & Qwen (Quinn)
Often called “China’s Amazon,” Alibaba’s AI arm has made Qwen the backbone of open-source LLM research.
Qwen’s Dominance
- Offers a full spectrum of model sizes: from phone-friendly 7B models to massive server-grade LMs
- Qwen 3 235B 2507 Reasoning is currently the #1 open-source model on benchmarks
- Released 100+ open-weight checkpoints since 2023 with **100M+ total downloads**
- Academic papers from top Chinese and U.S. universities now predominantly use Qwen
Multimodal Leadership
- Video generation model family: closed-source version ranks **7th globally**, open-source version is **#1 in open video gen**
ByteDance & ByteSeed
Though late to AI (strong presence only from mid-2024), ByteDance’s AI division ByteSeed exploded onto the scene.
Sudden Rise Through Investment
- Planned **$20 billion investment** in AI, including Chinese RAM and compute
- Launched **Doubao**, China’s **#2 AI assistant app** (as of March 2025), featuring:
- Text-to-speech
- Image & video generation
- Multimodal understanding
- Voice cloning
State-of-the-Art Generative Models
- Seed Dream 3.0: **#1 on Image Arena**, beating Google, Black Forest Labs, Flux
- Seed Think 1.5: Briefly held **SOTA Chinese reasoning model** title (though closed-source)
- Seed Video 1.0: **#1 on Video Arena**, outperforming legacy players “out of nowhere”
Tencent & HunYuan
Tencent entered AI cautiously but is now deeply embedded in China’s digital ecosystem.
Model Releases & Secrecy
- First LM announced: **HunYuan** (Sept 2023)
- First open-source model: **HunYuan Large** (late 2024)—briefly the **largest open model**
- Flagship reasoning model: **HunYuan T1** (minimal technical disclosure)
- Private hybrid model: **HunYuan Turbo S** (Mamba-Transformer hybrid) with a detailed paper
Multimodal & Ecosystem Integration
- Active in **open-weight video and 3D text-to-world models**
- Latest: **text-to-360° world model**
- Fully integrated into **WeChat** (owned by Tencent)
Level 3: Rising Chinese AI Startups
These agile companies are pushing boundaries with open models and niche expertise.
Moonshot AI
- Founded: **April 2023**
- Founder Yangzhi: ex-Meta AI & Google Brain; lead author of **Transformer-XL** (5,000+ citations) and **XLNet** (11,000+ citations)
- Raised **$200M at $300M valuation** within 2 months
- Launched **Kimi K2**, which briefly held the **non-reasoning open-weights crown**
- First AI mobile app to support **all 200,000 Chinese characters**
- Office culture: meeting rooms named after **iconic artists and albums**
Zhipu AI (ZAI)
- Originated from **Tsinghua University** (2019)—one of China’s earliest AI startups
- Flagship series: **GLM4.5** (hybrid reasoning model)
- Outperformed Kimi K2 with **30% fewer parameters**
- Ranked **top 3 on open-source LLM leaderboard**
- Released **ChatGLM** (June 2024): open-source chat model for English & Chinese with **41,000+ GitHub stars**
- Built the **first transformer-based text-to-video generator**
MiniMax
- Early mover: launched **Glow** (AI roleplay chatbot like Character.AI) in **late 2021**—**5M+ downloads**
- Also developed **Hyo AI**, a multimodal chat app
- Research output: **5 papers**, all with architectural breakdowns
Key Open-Source Releases
| Model | Release Date | Type | Notable Feature |
|---|---|---|---|
| MiniMax Text | Jan 2025 | Linear Attention LM | Efficient long-context processing |
| MiniMax VLM | Jan 2025 | Vision-Language Model | Strong vision understanding |
| MiniMax M1 | June 2025 | Hybrid Attention Reasoning | 1M context window; ranked #7 open LM |
| MiniMax M2 | July 2025 | State-of-the-Art | Beat GPT-4o OSS; #1 open-weights; #9 on AI Index |
Text-to-Speech Leadership
Contrary to assumptions that OpenAI and ElevenLabs dominate, MiniMax holds:
- #1 and #3 spots** on Artificial Analysis TTS Arena
“You should subscribe to BCloud. You should subscribe to BCloud. Can you tell the difference? Pretty good, right?”
Qwen (Quao) & Cling
- Known only for **Cling**, a private video generation model series
- Released during mid-2024 text-to-video wave
- Still **highly rated**, trailing only ByteSeed and Google
- No language model or research output—focused purely on video
- Parent company Quaiso operates similarly to TikTok
Level 4: Pre-DeepSeek Era Players
These companies peaked between GPT-4 Turbo and OpenAI’s o1, showing early promise but fading from prominence.
Baidu & ERNIE
- Launched **ERNIE 3.0** in **July 2021** (175B params, GPT-3 architecture)
- Remained **closed-source until July 2025**
- Released **ERNIE 4.5**: includes MoE vision-language models
- Not on major leaderboards—but worth testing
01.ai
- Founded by **Kai-Fu Lee**, ex-head of Google China
- November 2023: dropped a **near-SOTA open LM** with zero track record
- Briefly closed the gap with **GPT-4**, winning on key benchmarks
- May 2024: released private **Yi-Large**, led Chinese AI briefly
- Now **completely silent**—no news since early 2025
BRUN
- Released private models **Byron 3** (Jan 2024) and **Byron 4** (May 2024)
- Briefly led Chinese AI until **Qwen 2 series** dethroned them
- Now inactive—possibly unable to compete with Qwen’s pace
Level 5: The Underground Powerhouses
These labs fly under the radar but drive foundational progress.
StepFun
- First Chinese startup to build a **1T-parameter LLM** (Step 2, Nov 2024)
- Latest: **Step 3**—open-source multimodal reasoning model
- Vision understanding is **SOTA**, though language lags
- Only **4,400 downloads**—likely due to size and VLM complexity
OpenB (Open Lab for Big Model Base)
- Founded by **ModelBest Inc.**
- Uniquely publishes **both models and fine-tuned versions** of others’ models using their own datasets
- Flagship series: **MiniCPM & MiniCPMO**
- **MiniCPMO 2.6** rivals **GPT-4o** in language + vision (open-source)
- Innovative research: **RLPR** and non-technical-report papers
Huawei & Pangu: The Scandal
Huawei’s AI ambitions hit a major credibility crisis.
Whistleblower Revelations
- Noah’s Ark Lab (Huawei’s AI division) **wrapped rival open models** and rebranded them as Pangu:
- Qwen 1.5 10B → Pangu
- Qwen 2.5 14B → Pangu
- DeepSeek V3 → Pangu
- Added minor fine-tuning (few hundred billion tokens) and padded parameters to match advertised sizes
- Internal teams failed to train original models (38B, 135B, 718B) from scratch on Ascend chips
- A politically favored internal lab stole data, faked reports, and claimed bonuses
The GitHub post exposing this garnered **11,000+ stars**—a severe blow to Huawei’s AI credibility.
SenseTime & SenseNova
- Originated from **DeepID** (2014 facial recognition research)
- Now works on **autonomous vehicles, sensors, and medical AI**
- LM series: **SenseNova** since April 2023
- **SenseNova V6.5** (2025): claims to outperform **Gemini 2.5 Pro** and **Claude Sonnet**
- But: **no public access**—claims unverifiable
Shanghai AI Laboratory
- Founded by **Shanghai Municipal Government** (July 2020)
- Connects top talent from China’s elite universities
- Outputs **~20 top-tier AI papers per month**
- Known for **InternLM series**—small, efficient open models
- July 2025: released **Intern S1**—built on Qwen 3 235B + InternViT
- Matches **DeepSeek R1 0528** in language
- Adds **vision capabilities** unmatched by peers
Ant Group (Alibaba Affiliate)
- Behind **Alipay**; now deeply invested in AI research
- Pioneered concepts like:
- Large Language Diffusion Models
- Tag and Mingyi Tuni (well-received research)
- September 2025: released **Ring 1T**—a **1-trillion-parameter thinking model** comparable to SOTA open models
- Runs **Inclusion AI** initiative for all open-source releases
- Notable: Alibaba now has **two elite AI labs** (Alibaba Cloud + Ant Group)
Level 6: The Bridge Builders
These tools connect the fragmented AI world.
Cherry Studio by Qinghua Technology
- Open-source app (**32,000 GitHub stars**)
- Integrates **Chinese and international LLM providers**
- Available on **all operating systems**
- Features clean UI, native tools, and easy model switching
- Acts as a global bridge in an era of AI fragmentation
Level 7: The Deepest Tier – Unexpected Entrants
Companies with no AI history suddenly launching models.
Xiaohongshu (Red Note) & Retina High Lab
- Xiaohongshu: lifestyle platform (migration destination during U.S. TikTok ban)
- June 2025: launched **DOTS LM1** under **Retina High Lab**
- **142B total parameters, 14B active** (MoE architecture)
- Promising on **cost-performance scale**—details cut off in transcript but flagged as noteworthy
Honorable Mentions
Butterfly Effect & Manus AI
- Singapore-based (closed Chinese offices July 2025)
- Launched **Manus AI**: autonomous agent with **virtual computer environment**
- Can research, calculate, and browse the web independently
- Went viral in **March 2025**—even non-tech relatives heard of it
- Faced backlash: accused of **influencer-only access** and waitlisting others
- Company claimed **server scaling issues** (2M+ waitlist)
- Publicly available since **May 2025**
- Now faces competition as big labs enter the **AI agent application space**
Why the Chinese Iceberg Matters for Global AI
China’s AI ecosystem is not just copying—it’s innovating with distinct priorities:
- Open-source by default: DeepSeek, Qwen, Zhipu, and others release full stacks
- Hardware sovereignty: Native support for Chinese chips (e.g., DeepSeek V3.1)
- Multimodal focus: Video, 3D, TTS, and vision-language models lead globally
- Academic integration: Labs like Shanghai AI Lab feed university talent into real-world systems
Actionable Takeaways
- Track open-weight leaderboards: Chinese models dominate Hugging Face, Artificial Analysis, and LMSys
- Test MiniMax TTS: It may outperform ElevenLabs for your use case
- Use Cherry Studio: Simplify access to China’s best models
- Monitor Ant Group & Shanghai AI Lab: Their research often precedes industry trends
- Beware of inflated claims: Verify closed-source benchmarks (e.g., SenseTime, Huawei pre-scandal)
Final Thoughts: The Iceberg Is Melting—Fast
The Chinese AI iceberg is not static. What’s underground today (like Zhipu in 2019) becomes surface-level tomorrow. With U.S. policy potentially shifting toward restrictions, understanding this ecosystem isn’t just academic—it’s strategic.
As the transcript warns: “Before the US dude decides to make a diabolical move,” get informed. The open, collaborative, and fiercely competitive nature of China’s AI scene ensures that the next breakthrough could come from any layer—from a government lab in Shanghai to a startup in Shenzhen.
Stay curious, stay open-source, and keep watching the iceberg.
