French AI Company Mistral Launches Trillion-Parameter Open Source Model ML4
French company Mistral released a public preview of Mistral Large 4, abbreviated as ML4, codenamed le Chonk, on Tuesday. The company claims it is a native multimodal model with 1 trillion parameters, activating 49 billion parameters during inference, and is the largest model the company has developed to date. The preview interface is now available on Mistral Studio, with weights expected to be released by the end of this month.
The company focuses on code, agent processes, and multimodal understanding, claiming to excel in vertical tasks such as cybersecurity, manufacturing, and finance. WIRED's chief scientist Guillaume Lample stated that other labs will not invest the same level of attention in these areas. Mistral claims it can compete with the strongest open-source models globally and significantly surpass existing open-source weight models in the U.S. and Europe.
The training was completed at Mistral's proprietary data center in Europe, using 3,800 NVIDIA Grace Blackwell GPUs, with the public preview also provided by the same facility. VentureBeat reported that training started from scratch and took about two months. TechCrunch's VP of Technology Pierre Stock stated that the computing power is about one-half to one-third of that of Chinese counterparts, and significantly less than closed-source competitors. Before releasing the weights, safety tests will be conducted with the government and trusted partners, aiming for defensive use rather than malicious attacks.
The training data includes a large amount of multilingual content, covering over 160 languages, including all official EU languages. The model will be available in multiple regions globally, with European deployment operated end-to-end by Mistral, without relying on other digital service providers, and subject to European laws. The weights are expected to be licensed under Mistral's own terms, and the company has not yet disclosed the interface pricing.
In September, Mistral completed a €3 billion Series D funding round, with a post-money valuation exceeding €21 billion. Samsung Electronics led the round, with EQT-managed Scaleup Europe Fund and PSG Equity co-leading, and participation from a16z, NVIDIA, Salesforce Ventures, Advent, and BlackRock. The Series C was led by ASML. The funds will be used for expanding computing power, building infrastructure, and accelerating commercialization.
Buyers include enterprises, security officials, and governments seeking European deployment and open-source weights, while Mistral provides the preview using its own data center. This is a model preview, not the public release of weights. Computing expenditures remain in European data centers, benefiting the chip supply chain where Samsung and ASML are located. There will still be pressure on closed-source interfaces or European institutions relying on Chinese open-source weights.
Source: Public Information
ABAB AI Insight
Mistral was established in mid-2023 and has been positioned as Europe's counterpart to OpenAI and Anthropic. The valuation exceeding €21 billion was truly boosted by the €3 billion Series D led by Samsung in September. The Series C was led by ASML, which specializes in lithography, while the Series D was led by Samsung, which focuses on storage and chips. ML4 subsequently optimizes scenarios for chip design, manufacturing, and cybersecurity.
The 3,800 Grace Blackwell GPUs are housed in Mistral's proprietary data center in Europe, with the preview also utilizing the same facilities. The weights will only be released at the end of this month, after safety testing. Stock mentioned that this window is only for trusted partners and the government, locking in cybersecurity clients before the open-source release. The total parameters are 1 trillion, with only 49 billion activated, using a sparse structure to reduce inference costs.
The approach is similar to Meta's Llama making open-source weights the default for enterprises, and also resembles DeepMind's use of lower computing power to produce open-source models. The difference is that Mistral retains deployment jurisdiction under European law and includes Samsung and ASML as shareholders. The industry is currently in a phase of division between open-source weights and closed-source interfaces, with Europe seeking a third option that can be audited and run locally.
Essentially, this is a restructuring of the industry chain. Closed-source models charge by interface, while open-source weights charge by proprietary computing power. ML4 first occupies the government and security client space with the preview interface, then will deliver weights by the end of October. The financing is aimed at European data centers and chip shareholders, while the model offers an option that does not entrust critical tasks to U.S. closed-source or Chinese open-source solutions.
ABAB News · Cognitive Law
Weights released late, clients locked in first.
Total parameters are large, only activated ones count as costs.
Sovereignty is not just a slogan; it's about who controls the data centers.