World · Dec 2024 – Mar 2025
$5.6M dethrones the GPU-scale narrative
On December 26, 2024, V3 trained 671B parameters on 2,048 H800s for $5.6M and outperformed GPT-4o. A month later, R1 open-sourced reasoning via pure reinforcement learning at 1/27 the cost of OpenAI's o1. On January 27, the DeepSeek app topped the U.S. App Store, and Nvidia lost roughly $600B in market cap — the largest single-company daily drop in U.S. market history. The 'more compute wins' narrative was publicly shredded by a Chinese open-source team.
Events in this era
- December 26, 2024Read full →
$5.6M for GPT-4o-class performance
December 26, 2024: DeepSeek-V3 — 671B parameters, 37B active, trained on 14.8 trillion tokens. Just 2,048 H800 GPUs at a total training cost of about $5.6 million — less than one-tenth of comparable budgets. It outperformed GPT-4o and Claude 3.5 Sonnet across multiple evaluation sets.
- January 20, 2025Read full →
Reasoning awakens through pure RL
January 20, 2025: DeepSeek-R1 open-sourced under MIT license. The 671B model learned to reason entirely through reinforcement learning — no supervised fine-tuning cold start. It matched OpenAI's o1 on multiple reasoning benchmarks at one-twenty-seventh the inference cost. Six distilled versions shipped alongside, down to 1.5B parameters.
- January 27, 2025Read full →
App Store No. 1; Nvidia loses $600B in a day
January 27, 2025: DeepSeek's app overtook ChatGPT to top the U.S. App Store free chart. The same day, Nvidia's stock plunged — about $600 billion in market cap erased, the largest single-day drop for a single U.S. company in history. Investors began re-pricing the 'compute-pile = winning' AI arms-race narrative.
- March 24, 2025Read full →
A silent upgrade; math and coding past another line
March 24, 2025: DeepSeek-V3-0324 quietly updated — sharpened reasoning, frontend code, Chinese writing, function calling. No launch event, just a refreshed model card on HuggingFace and an MIT open-weight notice. Benchmarks showed it surpassing GPT-4.5 on math and coding.