Precision of completion suggester weights

Am I right in thinking that scoring based on completion suggester weights is only done to 24 bits of precision?

I am seeing that weights below 2^24 (ie weights below 16777216) are ranked correctly, but weights above this number that are close to each other often get the same score and hence are often ranked incorrectly. e.g. both 16777216 and 16777217 get a score of 16777216.

I am using Elasticsearch 2.3.2