Huge companies have dominated AI. Two startups with an entirely different approach think they can do better

Jaque Silva/NurPhoto/Getty Images

Among the AI heavyweights, Europe has long been missing a contender. The US has produced OpenAI, Google, Anthropic, Meta and SpaceXAI; while China has a whole host of AI labs. Now, one French startup is trying to best them all – and at a lower price.

On Tuesday, Paris-based Mistral unveiled its long-awaited model, ML4, which the company calls the “best open-weight model in the world.”

“Open-weight” refers to systems whose underlying workings are available for others to use and modify.

Mistral’s latest model is “on par” with the best-known closed models from the major AI companies in key areas, including “cyber defense, finance, (and) manufacturing tasks,” Mistral VP of Science Pierre Stock said in an interview.

Stock said Mistral’s new model was built at a “fraction of the compute” of competitors, making it cheaper for the company to build and for customers to use.

CNN was unable to immediately verify Mistral’s cost and capability claims.

Mistral is entering a tough market, even without including the trillion-dollar companies already dominating the AI industry.

Its “open-weight” approach is already dominated by Chinese companies, including DeepSeek, Alibaba and Z.ai . And the release of ML4 is also coming on the heels of another open-weight model release, Beam, from American startup Reflection.

Chang Martin/SIPA/AP

Reflection and Mistral’s performance, however, could help answer two questions hanging over the industry: whether open-weight models can compete with the best systems from companies like OpenAI and Anthropic; and whether Western companies can compete with China for the developers, businesses and governments that increasingly want AI that they can control themselves.

Mistral’s latest model was built and trained entirely within Europe, which CEO Arthur Mensch told CNN’s Becky Anderson is important as it means staying out of the China-US “duopoly.”

“There’s an enormous desire and an enormous need for alternative technology suppliers,” Mensch said.

Both releases are built around the same basic proposition: giving users more control over the technology.

Open-weight models such as Beam and ML4 can be downloaded and customized for specific tasks. The AI systems most people know today – OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini – are closed models, meaning their companies do not share the underlying “weights,” the billions of parameters that shape how the systems respond and act.

Open-weight models have generally been seen as lagging the capabilities of the best closed systems. They offer something those systems do not, however: the ability for users to modify the models themselves, often at a lower cost.

Martin Lelievre/AFP/Getty Images

That trade-off can matter when companies cannot send sensitive data to an outside provider. When AI company Hugging Face was hacked by rogue OpenAI agents that escaped a test environment, the company said it relied on an open-weight Chinese model to defend itself.

“What we realized during this attack was using a model on our own infrastructure, because all of that was private data, was mandatory and you can only do that with open models,” Hugging Face CEO Clement Delangue told CNN in July.

For open model AI companies, episodes like that are central to the pitch. They argue that many businesses do not need the biggest and most expensive models, but instead smaller, customizable systems that can run on their own infrastructure for security reasons. At the same time, they say the capability gap is narrowing while costs remain much lower.

While AI companies are famous for the amount of investor dollars poured into them, open-weight startups also have some serious backing.

Last month, Mistral announced it had raised more than $3.3 billion, which is more than half of the $6 billion the French company has raised since its founding three years ago.

That is still small compared with OpenAI and Anthropic, which have collectively raised more than $300 billion. But Mistral argues that being “starved in compute” because of its resource limitations forced the company to innovate.

Andrej Ivanov/AFP/Getty Images

Reflection, meanwhile, pairs a similar lower-cost pitch with deep-pocketed backers. The company is backed by Nvidia and major venture capital firms including 1789 Capital, where Donald Trump Jr. is a partner. Though it is only two years old, Reflection has raised billions of dollars and was among a small group of AI labs briefed by the White House this summer on the executive order directing the administration to review frontier models on a voluntary basis before they are released.

Like Mistral, Reflection said Beam delivers “strong model capabilities at lower cost,” calling it “a powerful workhorse model for enterprise coding and agentic workloads.”

Mistral and Reflection are not alone.

Other Western companies with open models include Cohere and Thinking Machines Labs, while larger AI companies have also offered open-weight models in the past. Meta chief AI officer Alexandr Wang said over the summer that Meta would soon launch new open models.

The push has also drawn support from the Trump administration, which has said it wants to bolster American open-weight models. Speaking on a panel with Reflection’s CEO at the Black Hat cybersecurity conference in August, National Cyber Director Sean Cairncross called open models “vital” for startups and innovation.

He said the government is “extremely interested in looking at ways to build US open source, make it competitive, make it the preferential adoption by planet Earth.”

雅克·席尔瓦/NurPhoto/Getty Images

在人工智能巨头云集的领域，欧洲长期以来一直缺少一位强有力的竞争者。美国拥有OpenAI、谷歌、Anthropic、Meta和SpaceXAI等公司；而中国也拥有众多人工智能实验室。如今，一家法国初创公司正试图以更低的价格超越它们。

周二，总部位于巴黎的 Mistral 公司发布了其期待已久的车型 ML4，该公司称其为“世界上最好的轻量级车型”。

“开放型”是指其底层工作原理可供他人使用和修改的系统。

Mistral 最新模型在关键领域（包括“网络防御、金融和制造任务”）与主要人工智能公司最著名的封闭模型“不相上下”，Mistral 科学副总裁 Pierre Stock 在一次采访中表示。

Stock 表示，Mistral 的新模型采用的计算能力仅为竞争对手的“一小部分”，这使得公司生产成本更低，客户使用成本也更低。

CNN目前无法立即核实“西北风”导弹的成本和性能声明。

即使不考虑那些已经主导人工智能行业的万亿美元公司，Mistral 也正进入一个竞争激烈的市场。

其“开放权重”模式目前已被包括DeepSeek、阿里巴巴和Z.ai在内的中国公司主导。此外，ML4的发布也紧随美国初创公司Reflection的另一款开放权重模型Beam之后。

张马丁/SIPA/美联社

然而，Reflection 和 Mistral 的表现或许有助于解答困扰业界的两个问题：开源模型能否与 OpenAI 和 Anthropic 等公司的最佳系统竞争；以及西方公司能否与中国竞争，以吸引那些越来越希望拥有自己可控人工智能的开发者、企业和政府。

Mistral 的最新型号完全在欧洲制造和培训，其首席执行官 Arthur Mensch 告诉 CNN 的 Becky Anderson，这很重要，因为这意味着要远离中美“双头垄断”。

门施表示：“市场对替代技术供应商有着巨大的渴望和需求。”

这两个版本都围绕着同一个基本理念：让用户对技术拥有更大的控制权。

像Beam和ML4这样的开放权重模型可以下载并针对特定任务进行定制。而如今大多数人熟知的AI系统——OpenAI的ChatGPT、Anthropic的Claude和谷歌的Gemini——都是封闭模型，这意味着它们的公司并不共享底层“权重”，也就是决定系统响应和行为方式的数十亿个参数。

人们普遍认为，开放式权重模型在功能上不如最好的封闭式系统。然而，它们提供了一些封闭式系统所不具备的优势：用户能够自行修改模型，而且成本通常更低。

Martin Lelievre/AFP/Getty Images

当企业无法将敏感数据发送给外部供应商时，这种权衡就显得尤为重要。人工智能公司 Hugging Face 曾遭到逃逸测试环境的 OpenAI 恶意代理攻击，该公司表示，它依靠一个开源的中国模型来保护自身安全。

“我们在这次攻击中意识到，使用我们自己的基础设施上的模型是必要的，因为所有这些都是私人数据，这是强制性的，而你只能使用开放模型才能做到这一点，”Hugging Face 首席执行官 Clement Delangue 在 7 月份告诉 CNN。

对于开放模型人工智能公司而言，这类案例是其宣传的核心。他们认为，许多企业并不需要规模最大、成本最高的模型，而是需要规模更小、可定制化程度更高、且出于安全考虑能够在自身基础设施上运行的系统。同时，他们指出，能力差距正在缩小，而成本却大幅降低。

虽然人工智能公司因获得大量投资者资金而闻名，但开放式初创公司也获得了相当可观的支持。

上个月，Mistral 宣布已筹集超过 33 亿美元，这超过了这家法国公司自三年前成立以来筹集的 60 亿美元的一半。

与OpenAI和Anthropic相比，这仍然微不足道，这两家公司累计融资超过3000亿美元。但Mistral认为，由于资源限制导致的“计算能力匮乏”迫使公司进行创新。

Andrej Ivanov/AFP/Getty Images

与此同时，Reflection公司则凭借类似的低成本方案和雄厚的资金支持，获得了英伟达和包括1789 Capital在内的多家大型风险投资公司的支持，而小唐纳德·特朗普正是1789 Capital的合伙人之一。尽管成立仅两年，Reflection公司已筹集了数十亿美元的资金，并且是今年夏天白宫向少数几家人工智能实验室通报情况的机构之一。白宫发布的行政命令指示政府在发布前沿模型之前对其进行自愿审查。

与 Mistral 一样，Reflection 表示 Beam 以更低的成本提供了“强大的模型功能”，称其为“企业编码和代理工作负载的强大主力模型”。

《米斯特拉尔》和《倒影》并非孤例。

其他拥有开放模型的西方公司包括Cohere和Thinking Machines Labs，而规模更大的AI公司过去也曾提供过开放权重模型。Meta首席AI官Alexandr Wang在今年夏天表示，Meta即将推出新的开放模型。

这一举措也得到了特朗普政府的支持，该政府表示希望加强美国的开放式创业模式。在8月份举行的黑帽网络安全大会上，国家网络安全总监肖恩·凯恩克罗斯与Reflection公司的首席执行官共同出席了一个小组讨论，他称开放式创业模式对初创企业和创新“至关重要”。

他说，政府“非常有兴趣研究如何发展美国的开源软件，使其具有竞争力，并使其成为全球首选的开源软件”。