DeepSeek's R1 model reached the top of the U.S. iOS App Store's free chart on Jan. 27, 20251. One day's surge helped erase roughly 18% of Nvidia's stock price. Eighteen months later, that tier had split three ways23. DeepSeek chases a 2027 IPO under founder Liang Wenfeng. Alibaba's Qwen reached 234 million app users, behind a team it keeps out of view. Mistral AI runs on the public face of chief executive Arthur Mensch. Open weights turned out to name a license, rather than a business model or a research culture.
Qwen's app reached 234 million users by May 2026, about a year after the Qwen3 family launched
Milestones for DeepSeek, Alibaba's Qwen and Mistral AI, January 2025 to July 2026, by month
Each pin is one milestone on a rail spaced by real time, colored by company, so pins standing at the same spot landed in the same month.
Sources: Wikipedia [1]; Wikipedia [2]; Wikipedia [3]; Mistral AI [4]. Chart by Kelly Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| Jan. 2025 | R1 tops the App Store. DeepSeek R1 tops the U.S. App Store; Jan. 27, 2025; the surge helped erase roughly 18% of Nvidia's stock price |
| April 2025 | Qwen3 launches. Qwen3 family launches; April 28, 2025, trained on 36 trillion tokens in 119 languages |
| Sept. 2025 | ASML leads €1.7B round. ASML leads Mistral's €1.7 billion round; At an €11.7 billion valuation |
| March 2026 | Mistral raises $830M. Mistral raises $830 million; For data centers near Paris and in Sweden |
| April 2026 | DeepSeek previews V4. 284-billion and 1.6-trillion-parameter builds |
| May 2026 | Qwen app: 234M users. Qwen app reports 234 million users; Hugging Face hosts more than 200,000 Qwen variants |
| May 2026 | DeepSeek $7B Series A. DeepSeek closes a $7 billion Series A; At a $52 billion valuation |
| July 2026 | DeepSeek readies IPO. DeepSeek confirms IPO preparations; For a possible 2027 listing |
A Quant Fund's Frontier Bet
Liang Wenfeng founded the hedge fund High-Flyer in 20151. He used its chips and cash to start DeepSeek, holding 84% through shell firms as of mid-2024. That history matters. DeepSeek's models came out of a trading firm's compute budget, rather than a funded research lab, which sets it apart from almost every rival here.
Release pace picked up through 2025 and 2026, far faster than most Western labs1. December 2024 brought V3-Base and its chat sibling V3. A refreshed V3-0324 followed that March under an MIT license open enough for almost any business use. August let the model switch between slow step-by-step reasoning and fast answers. September fixed language glitches users had flagged. December added a reasoning variant, closing the year with four updates in twelve months.
A V4 preview arrived in April 2026, spanning 284-billion and 1.6-trillion-parameter builds, and it skipped the widely expected "R2" name entirely1. Read that naming choice as a signal. DeepSeek's internal plan diverged from what outside analysts had mapped. Capital followed capability. The company closed a $7 billion Series A in May 2026 at a $52 billion valuation, reportedly third-highest among private AI companies, then confirmed IPO preparations for a possible 2027 listing that July. Anthropic separately accused DeepSeek in February 2026 of using fraudulent accounts to harvest Claude outputs for training. That dispute shows how contested the line between open research and competitive extraction has become.
Qwen's Anonymous Scale
Alibaba's Qwen3 family launched April 28, 20252. Dense models run from 0.6 billion to 32 billion parameters, beside sparse builds at 30B-A3B and 235B-A22B. Training covered 36 trillion tokens in 119 languages. Apache 2.0 covers the open tier, and a closed Qwen3-Max above one trillion parameters stays behind Alibaba's API.
Scale followed fast. The Qwen app reported 234 million users by May 2026, and Hugging Face hosts more than 200,000 Qwen variants2. One lightweight release, Qwen3-VL-2B-Instruct, passed 18 million downloads by itself. Reporting credits Qwen3 to Alibaba the company, rather than to any named researcher. Compare that to the profiles Western labs get for figures like Mensch. Alibaba runs Qwen as a product line, and puts reach ahead of the hiring pull a famous researcher brings.
Mensch's Personality-Led Path
Mistral took the opposite route. Arthur Mensch worked at Google DeepMind, co-founded Mistral in April 2023, and remains its chief executive and public face3. ASML led a €1.7 billion round in September 2025 at a €11.7 billion valuation, and reporting put the Dutch chip-tool maker's stake near 11%4. Christophe Fouquet, ASML's chief executive, framed it as more than money. "ASML is proud to enter a strategic partnership with Mistral AI, and to be lead investor in this funding round." Mensch named the logic. "This investment brings together two technology leaders operating in the same value chain." Read the phrase he picked. Same value chain. The firm that builds the machines that print the chips now owns a piece of the firm that trains the models those chips run.
Mensch kept buying through 2026. Koyeb, a Paris infrastructure startup, joined in February, beside a new partnership with Accenture3. An $830 million raise in March paid for data centers near Paris and in Sweden. Emmi AI, an Austrian simulation firm, joined in May. Each move pushes Mistral past model releases, into hardware and specific industries. That path differs from DeepSeek's independence and from Qwen's reach alike.
A Crowded Middle Tier
Moonshot AI, founded by Yang Zhilin, extended the pattern2. Its Kimi line grew into a K3 generation that reporting called the second-largest and second-strongest open-weight model by mid-2026, behind DeepSeek's flagship. Zhipu, which trades abroad as Z.ai, and MiniMax fill out a Chinese open-weight group big enough that Western labs began treating it as a bloc.
OpenAI answered with gpt-oss, its first open-weight release since GPT-2, shipping gpt-oss-120b and gpt-oss-20b under Apache 2.0 in August 2025. Read the timing as defense. It followed months of Chinese labs taking developer attention. Nvidia ran a parallel play with Nemotron, open reference models tuned for post-training and distillation rather than chat. That makes the chipmaker an open-weight contributor while it profits from every lab's training runs, license aside. Meta's Llama line slowed through 2026, as the company folded model work into its new superintelligence group, a turn toward tighter control and away from the open stance Llama was built on.
What the Category Proves
Three ways of operating reached roughly the same capability. That marks open weights as a licensing choice, one that fits almost any corporate shape, rather than a single strategy. A trading-fund spinoff, a cloud giant's product group and a funded European startup all got near the frontier through the same license, on separate paths to revenue.
Usage matters more than any benchmark. Qwen's 234 million users and 200,000-plus Hugging Face variants show real adoption by developers and by the public, which beats matching a closed lab on a test2. DeepSeek's App Store run in January 2025 made the point harder, at a scale that moved Nvidia's market value in one session1. Open weights once read as an underdog's pitch. The numbers now force closed labs to answer.
Clem Delangue, Hugging Face's chief executive, pointed to that same spread as a safety case in July 20265. Hugging Face hosts the Qwen and DeepSeek variants the piece already counts, each a download any developer can run on their own terms. "The biggest risk in AI is concentration of power. The way you make the world safer, in my opinion, is by leveling up the playing fields and creating transparency on these models," Delangue said. Read the noun he reaches for: concentration, the structure his argument targets, ahead of any single flag or company.
Matt Garman, AWS's chief executive, challenged that framing at the company's reInvent analyst session in December 2025, aimed at the economics an open license hides6. Garman argued a lab that pays billions to train a model keeps a reason to meter who uses it, adoption numbers aside. "If I spent billions to build it, I wouldn't give it away," Garman said. Read the conditional he chose: Garman treats giving a model away as a decision a lab makes, separate from whether users ever arrive.
By the numbers
- Eighteen percent: the drop in Nvidia's stock price tied to DeepSeek's Jan. 27, 2025 App Store surge1.
- Qwen reported 234 million app users as of May 20262.
- A $52 billion valuation followed DeepSeek's $7 billion Series A, closed in May 20261.
- Training for the Qwen3 family covered 36 trillion tokens across 119 languages2.
- €11.7 billion: Mistral AI's post-money valuation after ASML led a €1.7 billion round in September 20254.
- More than 200,000 Qwen model variants sit on Hugging Face, per the sources reviewed2.
- One release, Qwen3-VL-2B-Instruct, recorded 18 million downloads on its own2.
What to watch
DeepSeek's 2027 IPO prep tests one thing1. Does an open research group keep its license stance once reporting duties and shareholders arrive? Qwen's next release, set against its 234 million users, will show whether an unnamed team keeps winning reach, or eventually needs named researchers to hold credit2. Mistral's data centers near Paris and in Sweden, once running, will show whether Mensch's buying turns into capacity Mistral owns, rather than rents3.
Sources
- Wikipedia contributors, "DeepSeek," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/DeepSeek
- Wikipedia contributors, "Qwen," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/Qwen
- Wikipedia contributors, "Mistral AI," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/Mistral_AI
- Mistral AI, "Mistral AI raises 1.7B€ to accelerate technological progress with AI," Mistral AI, Sept. 9, 2025, https://mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai/
- Rebecca Bellan, "The real AI race may no longer be at the frontier," TechCrunch, July 14, 2026, https://techcrunch.com/2026/07/14/the-real-ai-race-may-no-longer-be-at-the-frontier-open-models-hugging-face/
- Stephen O'Grady, "How to Think About Open Weight Models," RedMonk, Sept. 3, 2026, https://redmonk.com/sogrady/2026/09/03/open-weight-models/