ARTEMIS LAB / BLOG
Mistral Large 4: choosing where your AI runs
Mistral releases Large 4’s weights. What that means for your data, the hardware required, and the checks to make before adopting it.
A business may want AI to work on its documents while retaining a choice of who runs it. On October 29, Mistral Large 4 becomes available with open weights on Hugging Face. Today’s news is the release of those weights, following a public preview earlier this month. Mistral’s announcement and Hugging Face’s release post.
What are weights?
Weights are the numerical settings a model learns during training. Downloading them lets you run your own copy with suitable software and hardware. Think of a machine whose engine you can install in your own workshop: you choose its location, but you also have to keep it running.
According to its official model card, Large 4 accepts text and images and produces text. A team could give it a screenshot with a question, or ask it to explain a document. These are possible uses to check against your own examples.
The card also describes a specific license, with a separate commercial license for businesses operating a model-as-a-service offering. Accessible weights still come with conditions. Check that point before building a commercial offering around the model. License information in the model card.
A workshop that needs room
Hardware matters. Hugging Face presents a deployment configuration using eight NVIDIA H200 graphics processors. The team also describes running heavily compressed versions on an Apple Silicon Ultra computer with 512 GB of unified memory. These are specific configurations, rather than a promise that the model will run on your laptop. Deployment details.
For a small business, our reading is straightforward: compare hardware, hosting, and maintenance costs with those of a managed service. Ask who will monitor failures and install updates, too. Downloading the model is only the beginning of the work involved in running it yourself.
Check performance on your own work
Mistral highlights improvements in coding and defensive cybersecurity since the preview. Those claims come from its announcement; we have not independently tested the model. They suggest areas to investigate, without establishing what the model will get right in your business. Results presented by Mistral.
To evaluate it, prepare a small fictional folder containing a table, a screenshot, and a deliberately missing piece of information. Ask for an explanation, then check each answer. Pay particular attention to whether the AI acknowledges what it does not know and whether a person can trace its claims back to their sources.
Choose who takes responsibility, too
Hosting the model yourself can help you control the route your documents take. That is a deployment possibility, rather than an automatic privacy guarantee: access permissions, logs, and connected tools still need examining. Draw the path a file follows, from upload through to deletion.
Before deciding, name someone responsible for access and someone who can check the answers. Control becomes practical when a person knows where the data goes and who can step in.