How can I change the 7-bit precision limit, it is too small?

Unfortunately, the precision is a critical limit in FHE and can’t be changed: we already have given to our users the maximum available precision at this time. What you can do is use quantization, post-training, and quantization-aware methods to make sure that the intermediate values in your model are compatible with the current limit.

You can be sure our research teams are working hard to increase this limit. We already have a lot of ideas that we need to implement to have a larger precision. Plus, we’ll have HW accelerations down the line, to make execution time less of a problem.

In the meantime, we are eager to see what you can already do with the current version of Concrete ML!

Today, it is 8b for Concrete-ML. It’ll be 16b next quarter