DeepSeek-R1 Released — a Reasoning Model Rivaling OpenAI o1 Shakes Global Markets
DeepSeek released a new reasoning-focused model, “DeepSeek-R1.” By applying large-scale reinforcement learning as a post-training step, the company said it achieved performance rivaling OpenAI’s o1 on benchmarks for math, coding, and logical reasoning — and released it fully open source under the MIT license, which became a major talking point worldwide.
Details
- Performance: Scored 79.8% on AIME 2024 (versus o1’s 79.2%) and 97.3% on MATH-500 (versus o1’s 96.4%), claiming reasoning performance on par with o1
- License: Released under the MIT license, allowing free commercial use and redistribution of distilled models
- Derivative models: Simultaneously released six distilled models ranging from 32B to 70B, claimed to match the performance of OpenAI’s o1-mini
- API pricing: Priced at $0.14 per million input tokens (with a cache hit) and $2.19 per million output tokens — remarkably cheap
- Rapid adoption: A week after release, on January 27, it overtook ChatGPT to reach #1 on the U.S. App Store’s free app rankings
What happened next
DeepSeek-R1’s arrival, dubbed the “DeepSeek shock,” directly rattled global tech stocks — NVIDIA shares dropped as much as 17-18% at one point on January 27. OpenAI’s Sam Altman even commented that it was “an impressive model, particularly around what they’re able to deliver for the price.” The event is remembered as one that instilled real alarm within the U.S. AI industry.