Nabugu - stock.adobe.com
Microsoft, AWS and Cerebras launch DeepSeek-R1 model
Tech companies are making DeepSeek available on their platforms while the cost-effective Chinese AI reasoning model has its 'viral moment.'
Enterprises that want to test DeepSeek-R1, the Chinese reasoning model that caused a tsunami in the tech industry, can get it from cloud providers AWS and Microsoft, the online platform GitHub and hardware maker Cerebras Systems.
The tech companies made the model from Hangzhou, China-based AI startup DeepSeek available this week, just days after it rocked the tech industry by casting doubt on the high cost of running AI in the United States. The model is comparable to OpenAI's o1 but requires a fraction of the GPU power, according to its developers.
"DeepSeek-R1 is certainly having its viral moment now," said Gartner analyst Arun Chandrasekaran.
Chandrasekaran said he expects many companies to offer DeepSeek-R1, including variants for specific industries and cloud, data center and edge deployments.
Model providers will likely differentiate themselves from competitors by offering better performance for the price through infrastructure innovations, Chandrasekaran said. They will also provide security and privacy layers and guarantees around legal indemnification.
"Having said that, in a few months, we may not remember R1 as much as we do today," Chandrasekaran said. "There is now a race to build models with better efficiencies, and we will see more such models from large cloud providers in the U.S., China, as well as from AI research labs in the world."
Microsoft made DeepSeek-R1 available on GitHub and the model catalog on Azure AI Foundry. The Foundry offers developers tools to experiment with, iterate on and integrate the model's capabilities into workflows. It also provides security and model evaluation tools.
AWS offers its version of DeepSeek on its SageMaker platform for building, training and deploying custom models. AWS customers can train DeepSeek on SageMaker using the Hugging Face open source platform.
"We expect to see many more models like this -- both large and small, proprietary and open source -- excel at different tasks," AWS said in an emailed statement.
Cerebras AI system with DeepSeek
Cerebras Systems introduced AI hardware running a 70-billion-parameter DeepSeek-R1 powered by the company's WSE-2 processor. The preview technology is contained in Cerebras' flagship CS-3 system, a unit that fits within a standard data center rack.
Cerebras claims DeepSeek-R1-70B is faster, more accurate, and less expensive than OpenAI's o1 reasoning model. Cerebras trained its model on the same data used to train Llama 3.3 using knowledge distillation, a technique that transfers data from a large, complex model to a smaller, more efficient one.
The price of the DeepSeek system will be similar to Cerebras hardware running Llama 3.3 70B. Cerebras has developed a framework called Cerebras Planning and Optimization to enhance the reasoning abilities of Llama 3.3
Cerebras declined to provide pricing details or say how many of its WSE-2 processors are used to run Llama and DeepSeek.
Hagay Lupesko, vice president of software engineering at Cerebras, said the company has received many queries from U.S. companies interested in its DeepSeek model.
"We have a pretty long list of preview customers who are in line for access to the model," Lupesko said. He declined to name any organizations but said many were in the consumer industry and interested in using the model for application development.
Cerebras also used the Llama 3.3 weights to train its version of DeepSeek. Weights are the numerical patterns critical to a model's ability to recognize patterns and make predictions based on input data. That means its DeepSeek model doesn't have the output restrictions of the original Chinese model, which, for example, does not answer queries about the Chinese leader, President Xi Jinping.
DeepSeek developed and released its model to the open source community. However, some components are inaccessible, and the company did not release the data used to train the model. Hugging Face and Llama creator Meta are attempting to reverse engineer the technology.
The company's researchers shook the tech world after they published a research paper demonstrating how they trained the model on significantly less-powerful Nvidia GPUs than the ones used for Open AI's reasoning model o1, yet achieved similar results. The disclosure raised serious questions about the efficiency of U.S.-made models and whether spending tens of billions of dollars on Nvidia's top chips was necessary.
On Monday, Nvidia lost $589 billion in market value due to the research paper and release of DeepSeek's R1 model. The company's stock has regained some, though not all, of its losses.
Antone Gonsalves is an editor at large for Informa TechTarget, reporting on industry trends critical to enterprise tech buyers. He has worked in tech journalism for 25 years and is based in San Francisco. Have a news tip? Please drop him an email.
