Jack Ma-backed Ant touts AI breakthrough on Chinese chips
" />
Ant Group, backed by Jack MA.
People said that ANT is using local chips, including from the alibaba Group Ltd. And Huawei Technologies, to train models using the so -called mixture of expert learning approach. They said he had results similar to those in NVIDIA Corp.
One person said that Hangzhou is still using NVIDIA to develop artificial intelligence, but it now depends on alternatives, including from Advanced Micro Devices Inc. And Chinese chips of their latest models.
Ant’s entry models have a race between Chinese and American companies that have been accelerated since Deepseek showed how able models can be trained for much less than billions invested by Openai and Alphabet Inc. It confirms how Chinese companies try to use local alternatives to the most advanced semiconductors in NVIDIA. Although it is not the most advanced, H800 is a relatively strong processor and currently prevents the United States from China.
The company published a research paper this month that its models sometimes excelled over Meta Platforms Inc. In certain criteria, which are not independently verified. But if you work as announced, ANT can determine another step forward to develop Chinese artificial intelligence by reducing the cost of conclusion or supporting artificial intelligence services.
When companies pour great money in artificial intelligence, MEE models appeared as a common option, gained confession to their use by Google and Hangzhou Startup Deepseek, among other things. This technology divides tasks into smaller groups of data, such as the presence of a of specialists, each of whom focuses on a slice of jobs, which makes the process more efficient. Ants refused to comment in an email statement.
However, the MEE models training usually depends on high -performance chips such as the graphics processing units sold by NVIDIA. The cost has been a ban so far for many small companies and a wider limited adoption. Ant works on ways to train LLMS more efficiently and eliminate this restriction. Its leafy title shows this, as the company determines the target to expand the scope of the “without excellent graphics processing units”.
This contradicts Nafidia. JenSen Huang has argued that the demand for account will grow even with the emergence of more efficient models such as Deepseek’s R1, which assumes that companies will need better chips to generate more revenues, not the cheapest models to reduce costs. It is stuck to the strategy of building large graphics processing units with more treatment, transistors and increased memory capacity.
Ant said it costs about 6.35 million yuan ($ 880,000) to train 1 trillion symbol using high -performance devices, but its improved approach will be reduced to 5.1 million yuan using low discrimination devices. The symbols are the information units that the model accommodates in order to recognize the world and provide useful responses to the user information.
The people said that the company plans to take advantage of the recent penetration in the large languages models that it developed, Ling-Plus and Ling-Lite, for artificial intelligence solutions, including health care and financing.
ANT has bought the Chinese online platform Haodf.com this year to enhance artificial intelligence services in health care. The company said in a separate statement on Monday that ANT has created AI Assist Asistant to support Haodf Doctors 290,000 with tasks such as medical record management.
The company also has the “Life Assistant” application of artificial intelligence called zhixiaobao and AI’s financial advisory service.
Regarding the understanding of the English language, Ant said in its paper that the Ling-Lite model was better in a major standard compared to one of the Llama models in Meta. Ling-Lite and Ling-Plus have surpassed the equivalent of Deepseek standards on Chinese standards.
Robin Yu, chief technology official at AI Solution Co. said.
Ant has made Ling Open Source models. Ling-Lite contains 16.8 billion of the teacher, which is adjustable settings that work like handles and faces to direct the performance of the model. Ling-Plus has 290 billion teachers, which are relatively large in the world of language models. For comparison, experts estimate that GPT-4.5 of Chatgpt has 1.8 trillion parameters, according to the Massachusetts Institute technology review. Deepsek-R1 has 671 billion.
The company has faced challenges in some areas of training, including stability. He said in the paper that small changes in the device or the structure of the model have led to problems, including jumps in the models error rate.
Ant said on Monday that she had built large model machines that focused on health care, which were used by seven hospitals and health care providers in cities including Beijing and Shanghai. She said that the large model benefits from Deepseek R1, QWEN’s QWEN and Ant and can implement medical consultations.
The company also said that it has launched two medical intelligence factors – Angel, who served more than 1,000 medical facilities, and a tribe, which supports medical insurance services. Last September, Manager Ai Healthcare was launched inside Alipay, its payment application.
This story was originally shown on Fortune.com
2025-03-24 08:15:00
