Jan 22, 2025
Open-source DeepSeek-R1
Open-source DeepSeek-R1
Open-source DeepSeek-R1
DeepSeek, a Chinese AI startup 🇨🇳, has released DeepSeek-R1, an open-source reasoning language model that matches the performance of OpenAI's o1 model across math, coding, and reasoning tasks. The model is available at 90-95% lower cost than o1 and is accessible through both open-source releases and an API.
DeepSeek, a Chinese AI startup 🇨🇳, has released DeepSeek-R1, an open-source reasoning language model that matches the performance of OpenAI's o1 model across math, coding, and reasoning tasks. The model is available at 90-95% lower cost than o1 and is accessible through both open-source releases and an API.
The Details
DeepSeek-R1 builds on the company's DeepSeek V3 mixture-of-experts model, achieving comparable scores to o1 on key benchmarks: 79.8% on AIME 2024, 97.3% on MATH-500, and a 2,029 Codeforces rating (better than 96.3% of human programmers).
The model was developed through a two-stage process, starting with DeepSeek-R1-Zero trained purely through reinforcement learning, followed by DeepSeek-R1 which combines supervised learning and RL to improve output quality and readability.
The pricing structure makes it significantly more accessible than o1, with input tokens costing $0.55 per million (vs o1's $15) and output tokens at $2.19 per million (vs o1's $60). The model is available through the DeepSeek chat platform as "DeepThink" and via API integration.
Why It Matters
Open-source AI has reached a major milestone. DeepSeek's release proves that free, public models can now deliver similar performance as leading closed models. This puts pressure on companies like OpenAI, as customers now have access to equally capable alternatives at a much lower price point.
The Details
DeepSeek-R1 builds on the company's DeepSeek V3 mixture-of-experts model, achieving comparable scores to o1 on key benchmarks: 79.8% on AIME 2024, 97.3% on MATH-500, and a 2,029 Codeforces rating (better than 96.3% of human programmers).
The model was developed through a two-stage process, starting with DeepSeek-R1-Zero trained purely through reinforcement learning, followed by DeepSeek-R1 which combines supervised learning and RL to improve output quality and readability.
The pricing structure makes it significantly more accessible than o1, with input tokens costing $0.55 per million (vs o1's $15) and output tokens at $2.19 per million (vs o1's $60). The model is available through the DeepSeek chat platform as "DeepThink" and via API integration.
Why It Matters
Open-source AI has reached a major milestone. DeepSeek's release proves that free, public models can now deliver similar performance as leading closed models. This puts pressure on companies like OpenAI, as customers now have access to equally capable alternatives at a much lower price point.
