Overview
VTP-Base-f16d64 is the Base checkpoint of MiniMax’s VTP visual-tokenizer research project. It learns visual representations through joint image–text contrastive, self-supervised and reconstruction training. The configuration specifies 16-pixel vision patches, a 64-dimensional visual bottleneck and a 256 × 256 image size. (github.com)
Capabilities and uses
- Extract image features for downstream vision tasks and support image–text similarity and zero-shot classification.
- Encode images into reconstruction latents and decode them back into images.
- Provide visual latents for separately trained diffusion-transformer image-generation systems. The tokenizer alone is not a complete text-to-image generator. (github.com)
Reported evaluation
The official Base model card reports 73.2% zero-shot accuracy, 81.0% linear-probing accuracy and 0.74 reconstruction FID. For generation, it reports 3.88 FID-50K using LightningDiT-XL trained for 80 epochs without classifier-free guidance. These figures describe the stated evaluation setup, not guaranteed application performance. (huggingface.co)
Limitations and license
- The project is described as work in progress. Its published examples and benchmarks do not establish performance across all image types or deployment settings. (github.com)
- The model card labels the license as modified MIT. The repository license requires qualifying commercial products or services exceeding either 100 million monthly active users or USD 30 million in annual recurring revenue to prominently display “MiniMax VTP”; review the full license before use. (huggingface.co)
Release date
December 16, 2025: the official project announced the technical report and pretrained weights. (github.com)
Pricing
As of November 17, 2026, no model-specific official API input or output price could be verified in either CNY/RMB on MiniMax’s mainland pricing page or USD on its international pricing page. Billing units and API availability for this checkpoint are likewise unverified; prices for other MiniMax models should not be applied to it. The pretrained checkpoint is available for download, but self-hosting may incur infrastructure costs. (platform.minimax.cn)
Official sources
- Project announcement and documentation
- Base checkpoint and model card
- Model configuration
- Project license
- Pricing checked November 17, 2026: mainland China pricing and international pricing.