The public weights change the deployment choice
Mistral has released downloadable weights for Large 4 under a custom license, according to VentureBeat’s report. That gives technically equipped organizations another option for operating and adapting the model themselves, alongside hosted access.
The report describes the final release as an efficiency improvement more than a broad jump in overall capability. In Artificial Analysis’s October 29 assessment, Large 4 at high reasoning scores 26 on Intelligence Index v5, compared with 25 for the preview. Its lead among available-weight models developed outside China is a narrower claim than worldwide open-weight leadership.
A task-cost result with a defined comparison
VentureBeat reports a fall in Intelligence Index task cost from $5.54 to $2.06, about 63%. That comparison is with Large 4’s preview in the stated evaluation. It does not establish the cheapest model for every workload or the full cost of operating downloaded weights.
The article also discusses a strong coding-agent result. Such comparisons combine a model with the software directing its tool use, so they should not be read as isolating model quality with all other components held constant. Different evaluation settings can change both rankings and resource use.
Match control and cost to the actual task
For an organization considering the weights, control over deployment is a separate reason to investigate the release. It still comes with infrastructure, integration and review work. The custom license also deserves direct examination before selecting a deployment arrangement; downloadable weights should not be casually described as unrestricted use.
A useful pilot would measure work completed to a defined standard, including correction attempts and operational overhead. The reported efficiency gain is a concrete comparison worth testing, while the small index-score change and geographically qualified ranking help keep expectations proportionate.