Four thousand processors challenge the giants, how Lu Shong tests the dominance of open models

French company Mistral AI launched its new multimodal model Mistral Large 4, internally nicknamed “Lu Shong”, in a move aimed at competing with US closed models and their Chinese rivals, based on what French President Emmanuel Macron described as the third path in artificial intelligence. The launch comes amid a growing split in the tech market between closed models that owners can cut off access to at any moment, and open-source models, most of which come from developers in China, as the company seeks to offer an alternative that breaks this binary.

The new model carries a trillion parameters and is currently available through a public endpoint governed by safety restrictions, while the company plans to release its open weights within three weeks after completing safety tests. Pierre Stoc, vice-president of the science division, explained that work is ongoing with trusted partners and governments to ensure the weights are used for technical protection and to prevent their misuse in malicious attacks, noting that an open model makes security auditing easier for companies and institutions that seek precise control over their software stack.

The most striking engineering paradox in training this model lies in hardware efficiency; the company trained it entirely on its own computing resources using only 4,000 Nvidia GPUs, a figure roughly half to a third of the hardware used by Chinese competitors and far less than that consumed by competing closed models.Only four thousand GPUs managed the training of a trillion-parameter modelwhich reflects a French wager on achieving record results through engineering focus and reduced computational waste before announcing official benchmark comparison results.

The company aims to employ the model’s capabilities in cybersecurity, financial services, and electronic chip design, a core activity of its main backers, including the Dutch firm ASML, which led its third investment round, and Samsung, which led the fourth round last month with a valuation of €21 billion, roughly $24.39 billion. With this step the company reaffirms its position as a laboratory leading frontier research, rejecting any shift that would be limited to merely providing inference services for other models.

This shift directly signals to those driving tech transformation in the Gulf, Egypt, and the Levant, as it offers a technical option for regional organisations torn between the risks of relying on closed Western clouds and security concerns about adopting Chinese software. Providing a model of this size with open, auditable weights that can be installed locally, and training it on limited hardware efficiency, enables banking sectors and regional governmental bodies to build sovereign computing systems and test financial-security scenarios without leaking data beyond national borders, while also reducing the server-farm requirements needed to fine-tune models locally.

The real turning point will begin as soon as the weights are made publicly auditable after three weeks.Then it will become clear whether the model can deliver on its performance promises in sensitive applications, or whether efficiency engineering will meet its practical limits in the face of the abundant computing power enjoyed by the giants of the East and the West.