Mistral has made Large 4’s trained weights downloadable, letting organizations run the model themselves, but the release does not make frontier AI cheap, fully transparent or risk-free.
How the weights arrived
The release followed a two-stage launch. On October 6, Mistral opened a public preview through its API and said the weights would follow by the end of the month. It held back the downloadable version while testing the model with cybersecurity firms, government agencies and other partners. On October 29, it published an improved version, saying three more weeks of reinforcement learning and user feedback had improved coding and defensive-security performance. Those are Mistral’s claims, not a guarantee the results will carry over to every organization’s tasks.
What open weights let users do
Open weights are the numerical values a model learns during training. Developers can use them to run Large 4 without sending every request to Mistral. The model accepts text and images and can process up to one million tokens of context. Its design uses only part of its network for each token. Mistral’s documentation lists 1.05 trillion total parameters and 52 billion active parameters; Hugging Face lists 49 billion active before counting embeddings and output layers. Organizations can host the model on their own systems, keep sensitive prompts on-site and set deployment rules to suit their needs.
The release does not include everything needed to recreate the model from scratch. Mistral published a summary of training content, but not the full training data and process. The license also sets limits: Hugging Face describes it as based on the MIT license but requiring a separate commercial license for model-as-a-service offerings. Self-hosting does not mean an organization can freely sell a competing hosted model.
Why running it takes major hardware
The trillion-parameter total does not tell the whole story about hardware needs. Large 4 uses fewer parameters for each calculation than a model that uses all its parameters each time, but its larger pool of experts still needs to be available across the hardware. At one byte per parameter, storing 1.05 trillion parameters would take about 1.05 terabytes, before extra memory for running the system. Mistral and Hugging Face also list a lower-precision NVFP4 version that substantially reduces the weights’ storage needs. Either way, running the model requires many connected servers, not a typical desktop. Mistral says it trained Large 4 on 3,800 Nvidia Grace Blackwell GPUs in its European data centers; that describes the training resources, not what inference requires.
Control comes with cyber risks
Mistral presents the model as an alternative to closed AI services, where a provider controls access, rules and infrastructure, and to powerful open-weight models from China. The case for independence is especially strong in cybersecurity: security teams may need to investigate a vulnerability or respond to an incident even when a hosted service’s safety filters block the request. Mistral says local deployment can keep data inside an organization and prevent a provider from cutting off access. The release may also raise Mistral’s profile and attract customers to its API and enterprise tools. Users gain some control, while Mistral remains a commercial provider.
The same tools that help defenders reproduce and fix software vulnerabilities can also help attackers understand them. Mistral said preview testing showed the model refusing attempts to attack live targets, but acknowledged it could still increase the ability to carry out real attacks. Safeguards come with the downloaded weights, but later changes to Mistral’s managed API will not automatically update copies already downloaded. Organizations running those copies must manage monitoring, access controls and updates themselves.
What organizations must test next
Companies and public agencies will need to test whether Large 4’s reported coding and cyber-defense performance holds up on their own work, and whether they can afford the hardware and staff to run it. Mistral says it will keep revising safeguards for its API and future releases; operators will need to update downloaded copies separately. The test of whether the release spreads AI power is not how many people download it, but how many can use it safely and affordably. Organizations that cannot may still depend on a cloud provider, even with the weights available.