Modal documentation - Storing model weights on Modal (Volume vs Image)

Fetched 2026-10-04. Weights options are a Modal Volume ('a distributed file system, a "shared disk"'), a CloudBucketMount for weights already in S3 or similar, or baking weights into the Image. THE RECOMMENDATION AND ITS REASON, verbatim: 'We recommend storing model weights in a Modal Volume ... Performance is similar for the two methods. Volumes are more flexible. Images are rebuilt when their definition changes, starting from the changed layer, which increases reproducibility for some builds but leads to unnecessa[ry] ...' - i.e. image-baked weights are not faster to start, they are just harder to change. Volumes read as an ordinary filesystem; there is no special read API to implement. This is the direct answer to 'object storage vs local disk vs baked image': on a serverless platform the three options are volume, bucket mount, or image, and the platform's own docs say performance is similar for volume and image, so the choice is about change frequency and cost, not latency.