Models
February 24, 2025 · View on GitHub
All our models are released under a research-only RAIL Model License.
Downloading the pretrained models
All our models described in our tech report are released on our GitHub.
For VaViM-L and VaVAM-L we use the following script to convert the weights to tar files:
# Performs something akin to:
# tar czf - filename | split -b 1900MB - filename.tar.gz.part_
python scripts/handle_checkpoints.py \
--mode create \
--checkpoint_dir XXXX \
--outdir vavam_release \
--maxsize 1900MB
The weights are then chunked into several tar files, you can merge them using the following command:
- Download all tar files (for VaViM-L or VaVAM-L).
- Put them in a single folder (e.g.,
vavam_l_release_chunks). - Run the following command:
# Performs something akin to:
# cat filename.tar.gz.part_* > filename.tar.gz
# tar xzf filename.tar.gz
python scripts/handle_checkpoints.py \
--mode extract \
--checkpoint_dir vavam_l_release_chunks \
--outdir vavam_release
The other models are released as single torch pickle files (can be loaded with torch.load).
Available models
Main models
Here are the links for our main VaViM and VaVAM models:
| model | # of params |
VaViM | VaVAM |
|---|---|---|---|
| VaVAM-S | 185M + 21M | part 1 | part 1 |
| VaVAM-B | 318M + 38M | part 1 | part 1 |
| VaVAM-L | 1.2B + 150M | part 1, part 2, part 3 | part 1, part 2, part 3, part 4 |
VaViM only
We also release the different checkpoints that helped up compute our scaling laws. Here are the different VaViM models, with different sizes and trained on different amounts of data:
| model | # params (in M) |
# data (in ×103) |
pre-trained | fine-tuned |
|---|---|---|---|---|
| VaViM-S | 185 | 38 | part 1 | part 1 |
| VaViM-S | 185 | 77 | part 1 | part 1 |
| VaViM-S | 185 | 116 | part 1 | part 1 |
| VaViM-S | 185 | 139 | part 1 | part 1 |
| VaViM-B | 318 | 38 | part 1 | part 1 |
| VaViM-B | 318 | 77 | part 1 | part 1 |
| VaViM-B | 318 | 116 | part 1 | part 1 |
| VaViM-B | 318 | 139 | part 1 | part 1 |
| VaViM-L | 1200 | 139 | part 1, part 2, part 3 | part 1, part 2, part 3 |
VaVAM
We trained VaVAM models given the VaViM models. Here are the different VaVAM models, with their corresponding amount of pre-training data:
| model | # params | # data (in ×103) |
VaVAM |
|---|---|---|---|
| VaVAM-S | 185M + 21M | 38 | part 1 |
| VaVAM-S | 185M + 21M | 77 | part 1 |
| VaVAM-S | 185M + 21M | 116 | part 1 |
| VaVAM-S | 185M + 21M | 139 | part 1 |
| VaVAM-B | 318M + 38M | 38 | part 1 |
| VaVAM-B | 318M + 38M | 77 | part 1 |
| VaVAM-B | 318M + 38M | 116 | part 1 |
| VaVAM-B | 318M + 38M | 139 | part 1 |
| VaVAM-L | 1.2B + 150M | 139 | part 1, part 2, part 3, part 4 |