Latest news with #MistralSmall3
Yahoo
17-02-2025
- Business
- Yahoo
Mistral releases regional model focused on Arabic language and culture
The next frontier for large language models (LLMs), one of the key technologies underpinning the boom in generative AI tools, might be geographical. On Monday, Paris-based AI startup Mistral -- which is vying to rival the likes of U.S.-based Anthropic and OpenAI -- is releasing a model that's a bit different from its usual LLM. Named Mistral Saba, the new custom-trained model is designed to address a specific geography: Arabic speaking countries. The goal for Mistral Saba is to excel in Arabic interactions. Mistral Saba is a relatively small model with 24 billion parameters. As a reminder, fewer parameters generally leads to better performance with lower latency. But more parameters usually means smarter answers, even though it's not a linear correlation. Mistral Saba is comparable in size to Mistral Small 3, its general purpose small model. But, according to Mistral's own tests, Mistral Saba performs much better than Mistral Small 3 when handling Arabic content. As an interesting side effect, due to cultural cross-pollination between the Middle East and South Asia, Saba also works well with Indian-origin languages, per Mistral -- especially South Indian-origin languages, such as Tamil and Malayalam. The new model represents an interesting strategic move for the French AI giant, showing an increased focus on the Middle East. Mistral said it expects the model will help it gain traction among customers in the region. As an off-the-shelf model, Mistral Saba could be used for conversational support or content generation in Arabic that sounds more natural and relevant. It can also be used as the basis for some fine-tuned models for internal use cases, the company said. Last week, Mistral used the AI Action Summit to demonstrate that it's getting serious about business. While the company has already raised large amounts of money from international investors, many of its foreign backers are based in the U.S. -- investors such as Lightspeed Venture Partners, Andreessen Horowitz and Salesforce Ventures. Due to the shifting geopolitical landscape, Mistral could potentially welcome Middle Eastern investors in its upcoming funding round. It would be a way to raise more money to remain relevant in the AI race on a technical level, while positioning itself as the international alternative to U.S. and Chinese AI companies. Mistral's newest model, Saba, could therefore contribute to that potential fundraising effort. Mistral Saba is accessible through Mistral's API. It can also be deployed on-premise, which could be a strong selling point for companies working in sensitive industries, such as energy, finance or healthcare. Due to the company's European roots, since the release of the original open-weight Mistral 7B model it has often reiterated that it takes multi-language support seriously. Saba's release is a continuation of that positioning. And Mistral said that it will be turning its attention to other regional languages down the road.
Yahoo
07-02-2025
- Business
- Yahoo
As US and China dominate AI race, where is Europe's answer to DeepSeek and ChatGPT?
Europe has trailed the US and China in the AI race, and the rise of DeepSeek has only widened the gap, but open-source projects and regulatory efforts could help the continent carve its own path in artificial intelligence, according to tech experts. Hangzhou-based start-up DeepSeek made headlines last month with two large language models - V3 and R1 - that have emerged as challengers to OpenAI's ChatGPT while requiring only a fraction of the cost and computing power to build. This has put China in a strong position in its AI rivalry with the United States and fuelled hopes for more DeepSeek-style disrupters. Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team. But European tech firms have not yet produced an AI contender on the level of ChatGPT or DeepSeek. Neil Lawrence, a senior AI fellow at the London-based Alan Turing Institute, noted that Britain was "a long way behind" the US and China in both the development and deployment of the technology. "Similar applies to most of Europe, but there are encouraging signs in Germany, France, Finland and Switzerland," Lawrence said. DeepSeek's success stands in stark contrast to the French government-backed chatbot Lucie, which was deactivated on January 25 after it gave nonsensical answers to simple maths problems and even suggested a user consume "cow's eggs". Lucie was developed as part of the €54 billion (US$56.2 billion) France 2030 investment programme. It aimed to challenge the dominance of the English language in AI and offer an alternative to models such as ChatGPT. However, its underwhelming performance sparked criticism and online ridicule. Jeroen Groenewegen-Lau, who heads the science, technology and innovation programme at the Mercator Institute for China Studies, said it would be difficult for Europe to compete with key players such as China and the US. "Maybe it's not realistic to think that you will compete at that global cutting edge of general intelligence, but there's a lot of value in being a close follower and working on some of the implementation and focusing on some of the specific areas," he said. Despite the anxiety DeepSeek has sparked in Europe, open-source models have offered a glimmer of hope. On January 30, Paris-based start-up Mistral AI unveiled a new open-source language model called Mistral Small 3. The company said Mistral Small 3 rivalled the performance of models three times its size while significantly cutting computing costs, positioning it as an "excellent open replacement" for GPT-4o mini - OpenAI's scaled down, cheaper model - and a complement to DeepSeek. On Monday, an alliance of 20 European research institutions announced OpenEuroLLM, an ambitious AI initiative aimed at developing a European version of DeepSeek. Backed by the European Commission, the open-source project has a budget of €52 million and could be the commission's biggest AI venture, according to tech news site The Next Web. A statement from OpenEuroLLM said the project aimed to fortify "Europe's competitiveness and digital sovereignty". Lawrence said that while Britain's current AI output was "flimsy", DeepSeek's success could help encourage change. "The UK's falter has mainly been due to poor and panicky policy advice, so we can hope it will be quickly reversed because the foundations are still strong," he said. Anthony Cohn, professor of automated reasoning at the University of Leeds, said there was some evidence that foundation model performance was plateauing as the release of GPT-5, OpenAI's next large AI model, had been delayed. Foundation models are AI models trained on vast data sets that can be applied to a wide range of tasks. "Assuming LLM [large language model] development is plateauing, then that gives a chance for others to catch up, especially since Western nations will have access to the latest Nvidia [chips] currently banned for export to China," Cohn said. "Moreover, if the development costs quoted by DeepSeek are real, then this certainly puts the development of similar models within reach of many nations." Cui Hongjian, head of European Union studies at Beijing Foreign Studies University, said Europe also aimed to maximise its advantages in regulation to carve out a different path to AI development. Last year, EU policymakers passed the Artificial Intelligence Act, marking one of the world's first comprehensive efforts to regulate the rapidly evolving technology. On Saturday, Britain became the first country to criminalise AI tools used to create sexually abusive images. "However, without the support of actual technological and industrial development, I believe these regulations cannot stand firm," Cui said. Cui cautioned that excessive regulation might stifle innovation, adding that there were signs that AI had become "politicised" in some European countries. "The impact of DeepSeek will continue to unfold, influencing Europe's ongoing discussions on how to enhance its competitiveness. The focus will likely shift towards creating regulations that foster innovation rather than hinder it," he said. "DeepSeek, as it may represent, continues to uphold a globalised and open landscape, which I believe aligns more closely with Europe's interests." This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice reporting on China and Asia for more than a century. For more SCMP stories, please explore the SCMP app or visit the SCMP's Facebook and Twitter pages. Copyright © 2025 South China Morning Post Publishers Ltd. All rights reserved. Copyright (c) 2025. South China Morning Post Publishers Ltd. All rights reserved.
Jordan Times
31-01-2025
- Business
- Jordan Times
France's Mistral hails DeepSeek's AI model
PARIS — French AI startup Mistral on Thursday hailed Chinese competitor DeepSeek's R1 model as 'great' for the fast-developing sector, while announcing another new release of its own. Mistral has in the past been hailed for offering models — the engines that power generative AIs — comparable with American giants at a fraction of the cost in energy and cash. DeepSeek claimed that its R1 release pulled off a similar feat, causing near panic in the US this week and battering tech stocks. R1 is 'a great and complementary piece of open-source technology', Mistral said in a statement on its website, referring to the fact that elements of DeepSeek's technology are available for other researchers to use and build on. The French firm's own Thursday release, 'Mistral Small 3', is 'competitive with larger models' including Meta's Llama or Qwen, developed by Chinese e-commerce giant Alibaba, Mistral said. It added that Small 3 was 'the most efficient model of its category' and might be used for tasks ranging from customer service to detecting financial fraud or controlling robots. The new model 'complements large open-source reasoning models like the recent releases of DeepSeek, and can serve as a strong base model for making reasoning capabilities emerge', Mistral said. Like DeepSeek, Mistral offers elements of its technology as open-source tools for other developers. The French underdog shot to prominence following its 2023 founding by Arthur Mensch, Guillaume Lample and Timothee Lacroix, becoming Europe's greatest hope of matching US heavyweights. Mistral took in 600 million euros (626 million) in a summer fundraising round, bringing its valuation to around 6 billion euros. AFP news agency signed a deal with Mistral in mid-January allowing the startup's chatbot to draw on the news agency's articles to formulate responses. Some in the US have accused DeepSeek of simply piggybacking on the work of American AI developers like OpenAI's ChatGPT models, rather than achieving its high performance independently. Page 2
