DeepSeek releases DeepSeek-R1, an open-source AI model outperforming OpenAI at lower cost
The Chinese AI startup DeepSeek has launched its new reasoning AI model, DeepSeek-R1, under an open MIT license. This model has surpassed OpenAI's o1 in several benchmarks for math, coding, and reasoning. With an impressive 671 billion parameters, R1 is accompanied by six smaller, distilled versions that range down to 1.5 billion parameters, enabling them to operate on local devices.
The DeepSeek-R1 family includes various models, such as DeepSeek-R1-Zero, which uses reinforcement learning without supervised fine-tuning, and six compact models with parameter sizes ranging from 1.5 billion to 70 billion. The model employs a mixture-of-experts structure, optimizing performance and cost efficiency.
Remarkably, DeepSeek-R1 was developed at a much lower cost than OpenAI’s o1, with API access starting at $0.14 per million tokens, significantly lower than OpenAI's $7.50 for the same tier. This makes R1 a cost-effective alternative, undercutting OpenAI’s unlimited o1 access priced at $2,400 annually through ChatGPT Pro. By releasing its models as open-source, DeepSeek aims to democratize access to advanced AI, and challenge established players like OpenAI with its cost-effective and high-performing models.
- AI Chatbot
- Freemium • Open Source
Related news
External links
- DeepSeek's Thread on X@deepseek_ai on X • Official source
- DeepSeek-R1 ReleaseDeepSeek API Docs • Official source
- DeepSeek's new open-source AI model can outperform o1 for a fraction of the costZDNET
- DeepSeek releases open-source reasoning model, DeepSeek-R1, beats OpenAI’s o1The Hindu
- DeepSeek Crushes OpenAI o1 with an MIT-Licensed Model—Developers Are Losing ItAIM Media House
Comments
Ah, capitalism's worst nightmare: a free and open-source AI model that actually works better than the overpriced, proprietary offerings from big tech.
It is remarkable how a firm can occasionally come out of seemingly nowhere and create such havoc. And I mean that in the best possible way. My greatest point of interest is on these distilled LLMs, because access to these allows home users the potential to have a private, but highly useful personal AI on their own affordable commodity hardware, reducing the need for expensive datacenters owned by wealthy corporations. We will need more testing to see whether these distilled models hold up, but if they are as promising as they first appear, the future just got a bit more interesting.
One of the weakness of OpenAI (excluding their hallucinating storyteller CEO) is its unproductive company environment (about 1.5k employees, paid up to $1M a year), it was only a matter of time, once the aura from this most popular AI-focused company in the world has faded out, that its major burden is to find enough cash (like its last $6B funding rising) twice a year to support their increasing cost. Many studies have asserted that open source models were much efficient (i.e. reduced cost per query) that proprietary ones (which is mostly related to open-source nature, just as any software), and given Chinese progress in processor architectures (Loongarch, and also RISC-V), they have enough computing power to become a major player even in this saturated AI field. As long as money pours into OpenAI pockets from fund raising, and that Microsoft keeps building insane datacenters to justify OpenAI existence, the roles will not be reversed so easily, but the industry is more and more pressured to make money from AI to cover the $100B+ a year investment in this technology, and companies are increasingly burning money to hope finding AGI or whatever thing, that LLM could never be able to offer.