Mistral 推出万亿参数新模型，瞄准超越闭源与开源竞争对手

开放与闭源 AI 之间的竞赛正在展开，欧洲仍在其中占有一席之地。周二，法国 AI 实验室 Mistral AI 发布了 Mistral Large 4（ML4），这是一款新的大型多模态模型，目标是超越美国和中国的竞争对手——这也呼应了法国总统马克龙所描述的“AI 第三条道路”。

在可被断开服务的闭源模型，与通常由中国公司开发的开放模型之间，分歧日益加深。Mistral 将 ML4 定位为两者之外的替代选择。ML4 因其 1 万亿参数而被昵称为“Le Chonk”；它显然并不小，但目前还不是一款开放权重模型。现阶段，它只能通过带公共安全护栏的端点访问；不过，Mistral 计划在完成安全测试后的三周内提供其权重。

“Mistral 科学副总裁 Pierre Stock 向 TechCrunch 表示：‘与此同时，我们将与可信赖的合作伙伴和政府合作，确保开源权重能够用于防御，而不是用于[实施]恶意攻击。’”

近几个月来，安全担忧持续增加，尤其是在 Mistral 的核心受众——企业和机构——之中。与此同时，Stock 表示，开放权重模型也更容易接受审计。

另一个重要的幕后因素是，ML4 完全使用 Mistral 自身的算力完成训练；仅使用了 4,000 块 Nvidia GPU。Stock 表示，这一数量“比我们的中国竞争对手少两到三倍，也显著低于闭源竞争对手”。

基准测试结果尚待公布。Stock 表示，Mistral 希望 ML4 成为开放权重模型中同类最佳的产品，尤其是在中国以外的市场，但并不局限于此。得益于针对性的训练，它还有望在某些对客户至关重要、且多模态能力能够创造价值的特定领域超越闭源模型。

Stock 表示，ML4 的优化应用场景包括网络安全和金融，也包括芯片设计；后者正是 Mistral 两大主要投资方的核心业务——荷兰巨头 ASML 曾领投其 C 轮融资，三星则在上个月以 210 亿欧元估值（约合 243.9 亿美元）领投其 D 轮融资。

当时，该公司曾试图传达一个信号：其托管中国模型的决定，并不意味着公司转向仅仅成为一家推理服务提供商。随着“Le Chonk”加入阵营，该公司相信自己仍应被视为一家前沿 AI 实验室。

The race between open and closed AI is on, and Europe is still in the mix. On Tuesday, French AI lab Mistral AI released Mistral Large 4 (ML4), a new large multimodal model aiming to leapfrog both American and Chinese rivals — following what French president Macron described as “a third way in AI.”

Amid a growing divide between closed models that can be unplugged and open models that are often made in China, Mistral is positioning ML4 as an alternative to both. Nicknamed Le Chonk in reference to its 1 trillion parameters, ML4 is definitely not small; but is not an open-weight model yet. For the time being, it can only be accessed via a public guardrail endpoint, but Mistral plans to make its weights available in just three weeks, after safety testing is complete.

“In the meantime, we’ll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks,” Mistral VP Science Pierre Stock told TechCrunch.

Security concerns have been mounting in recent months, particularly among Mistral’s core audience — enterprises and institutions. At the same time, an open-weight model is easier to audit, Stock said.

Another important behind-the-scenes aspect is that ML4 was trained entirely on Mistral’s compute; using only 4,000 Nvidia GPUs “which is two to three times less than our Chinese competitors, and significantly less than the closed source competitors,” Stock said.

With benchmark results still pending, Mistral hopes ML4 will be best in class among open-weight models, especially outside of China, but not only, Stock said. Thanks to focused training, it could also outperform closed models in specific areas that are key to its customers, and where multimodal capabilities can add value.

According to Stock, ML4’s optimized use cases include cybersecurity and finance, but also chip design, which is core to two of Mistral’s main backers — Dutch giant ASML, which led its Series C, and Samsung, which led its Series D last month at a €21 billion valuation (about $24.39 billion).

At the time, the company tried to convey that its decision to host Chinese models wasn’t a pivot into becoming a mere inference provider. With Le Chonk in its corner, the company believes it should still be considered a frontier lab as well.