Table of Contents
Deepseek
Founding as a Hedge Fund Spin-Off
DeepSeek was founded in July 2023 as a spin-off from the AI division of High-Flyer, a Chinese quantitative hedge fund based in Hangzhou. 1)
Founder's Academic Roots
Founder Liang Wenfeng was born in 1985 and studied machine vision at Zhejiang University, where he began researching the topic in 2008 before starting his company with two classmates. 2)
The GPU Stockpile Before Export Controls
Between 2019 and 2022, High-Flyer spent approximately $139 million and stockpiled 10,000 Nvidia A100 GPUs to build a supercomputer—before U.S. export restrictions tightened. 3)
A Tiny Team Against Giants
DeepSeek had roughly 200 employees when it released V3 and R1, compared with at least 3,500 at OpenAI, according to headcount estimates. 4)
V3's Arrival and Scale
DeepSeek released V3 via API on December 26, 2024. It carries 671 billion total parameters but activates only 37 billion per token, keeping inference costs closer to a much smaller model. 5)
R1's Reasoning Debut
The original DeepSeek-R1 was released on January 20, 2025, as a reasoning-focused model that generates an explicit reasoning trace before producing its final answer. 6)
One Checkpoint, Two Personalities
V3 and R1 share the same pretrained checkpoint and architecture, then diverge through different post-training pipelines—V3 for fast direct answers, R1 for visible multi-step reasoning. 7)
R1's Self-Correction Habit
R1 is trained to produce longer reasoning traces, self-correct mid-generation, and try different problem-solving strategies before committing to an answer. 8)
V3's MMLU Score
DeepSeek-V3 scored 88.5 on the MMLU benchmark, placing it in the same range as other frontier general-purpose models. 9)
Three Times Faster Than V2
DeepSeek-V3 generates at 60 tokens per second, roughly three times faster than its predecessor V2. 10)
V2's Architectural Turning Point
DeepSeek-V2, released in May 2024, had 236 billion total parameters with 21 billion activated per token and introduced Multi-head Latent Attention (MLA) and DeepSeekMoE. 11)
Four Innovations Behind the Hype
DeepSeek's technical papers highlight multi-head latent attention (MLA), mixture-of-experts (MoE), multi-token prediction (MTP), and group relative policy optimization (GRPO) as its core algorithmic contributions. 12)
R1-Zero Learned Reasoning Without Examples
DeepSeek-R1-Zero showed that reinforcement learning at scale could directly enhance reasoning capabilities without supervised fine-tuning, using GRPO with rule-based rewards. 13)
R1-Zero's Contest-Level Math
R1-Zero reached 71.0% pass@1 on AIME 2024, achieving contest-level math performance from a base model with no supervised fine-tuning. 14)
The $294,000 Training Claim
DeepSeek said its R1 model was trained for just $294,000 using 512 Nvidia H800 chips, far below what U.S. rivals spent. 15)
The $5.6 Million V3 Claim
DeepSeek claimed it trained its R1 chatbot on its V3 LLM for just $5.6 million, using less advanced Nvidia H800 chips than U.S. companies had access to. 16)
The 2,000-Chip Revelation
DeepSeek's engineers said they needed only about 2,000 Nvidia chips to train V3, while leading AI companies typically used supercomputers with as many as 16,000 chips or more. 17)
Matching the Best on a Fraction of the Budget
When DeepSeek introduced V3 the day after Christmas 2024, it matched the abilities of the best chatbots from U.S. companies like OpenAI and Google. 18)
The App That Started the Panic
DeepSeek released its first free chatbot app on January 10, 2025, based on the V3 model, setting the stage for the global reaction that followed. 19)
Nvidia's Record One-Day Loss
Nvidia lost nearly $600 billion in market value in a single day—the biggest drop in U.S. stock market history—after DeepSeek's R1 release shook assumptions about AI costs. 20)
"AI's Sputnik Moment"
Venture capitalist Marc Andreessen called DeepSeek's rise “AI's Sputnik moment,” capturing how a little-known Chinese startup had upended assumptions about who could build frontier AI. 21)
Topping the App Store in Two Countries
DeepSeek surged to the top of Apple's free app download charts in both the United States and China, surpassing ChatGPT and marking the first time a Chinese app held the No. 1 spot in both regions. 22)
Italy's Data Protection Block
Italy's data protection authority, Garante, blocked access to DeepSeek to protect users' data and opened an investigation after the company said European legislation did not apply to it. 23)
The U.S. Federal Device Ban
The U.S. Department of Homeland Security ordered federal civilian agencies to stop using DeepSeek-developed tools and remove related software from government IT devices. 24)
The MIT License Gamble
DeepSeek released R1's code and model weights under the MIT License, allowing anyone to distill, modify, and commercialize the model freely. 25)
A 685-Billion-Parameter Open Release
DeepSeek released V3.1-Terminus, a 685-billion-parameter model, under a permissive MIT License, enabling global developers to download, modify, host, and commercialize it. 26)
Coder Was Built for Repositories
DeepSeek-Coder was trained with a repository-level methodology, allowing it to understand and generate code across entire project structures rather than isolated snippets. 27)
Math Was a Dedicated Track
DeepSeekMath was developed as a specialized model for mathematical reasoning, using reinforcement learning breakthroughs later carried into the R1 line. 28)
Vision Came Through DeepSeek-VL
DeepSeek-VL and its successor Janus were decoupled vision-language frameworks designed for real-world multimodal understanding, from diagrams and formulas to web pages and natural images. 29)
Reasoning Behaviors Emerged, Not Engineered
During R1-Zero training, DeepSeek observed that powerful and interesting reasoning behaviors—long chains of thought, self-reflection, verification, and dynamic strategy adaptation—emerged naturally after thousands of reinforcement learning steps. 30)
GRPO's Full Name
GRPO stands for Group Relative Policy Optimization, the reinforcement learning framework DeepSeek used to train R1-Zero without a separate reward model. 31)
V4 Embraced Huawei Silicon
DeepSeek's V4 model highlighted compatibility with chips made by Chinese technology giant Huawei, a significant shift in the hardware stack. 32)
The World's Richest AI Founder
Liang Wenfeng's net worth reached $36 billion, making him the world's richest AI founder ahead of Anthropic's Dario Amodei and OpenAI's Greg Brockman. 33)
Liang's Unusual Ownership Stake
Liang retained control of nearly four-fifths—about 78%—of DeepSeek, an unusually concentrated ownership structure among AI founders. 34)
A Hangzhou Story
DeepSeek is based in Hangzhou, the same city where High-Flyer was founded in 2015, making it a rare AI frontier lab rooted outside Beijing and Shenzhen. 35)
No App-Building, Just Research
Under Liang's leadership, DeepSeek deliberately avoided app-building and concentrated research talent and resources on creating a model that could match or beat OpenAI. 36)
The Originality Gap
Liang said the real gap between China and the U.S. in AI is “the difference between originality and imitation,” warning that without change China would always be only a follower. 37)
First Model, Same Year
DeepSeek released its first AI large language model later in 2023, only months after the company was created in July of that year. 38)
R1-Zero's Minimalist Recipe
A follow-up study found that a minimalist R1-Zero-style recipe could achieve 43.3% accuracy on AIME 2024 with a 7B base model, establishing a new state of the art at that scale. 39)
Millions of Downloads in Days
Italy's regulator noted that the DeepSeek app had been downloaded by millions of people around the globe in just a few days before it was blocked. 40)
The High-Flyer Cash Cow
Liang's hedge fund, Zhejiang High-Flyer Asset Management, oversees more than 70 billion yuan ($10 billion) and generated an average return of 56.6% across its funds in 2025, funding DeepSeek's research. 41)
V3 Was Fully Open-Sourced
DeepSeek released V3's models and papers fully open-source alongside the API launch, making the architecture and weights available to the broader research community. 42)
The $7.4 Billion Raise
DeepSeek raised $7.4 billion in June 2026 at a $50 billion valuation, about five times the $10 billion valuation reported just two months earlier. 43)
Six Distilled R1 Variants
DeepSeek released six smaller distilled R1 variants ranging from 1.5B to 70B parameters, based on Qwen and Llama backbones, inheriting R1's reasoning patterns at lower compute cost. 44)
R1-0528 Added Tooling
The R1-0528 update on May 28, 2025, pushed reasoning quality higher and added function calling and JSON output support that the original R1 release lacked. 45)
The R1 Paper Landed on Inauguration Day
DeepSeek published its R1 research paper on the same day as Donald Trump's inauguration, a timing coincidence that amplified the market shock that followed. 46)