← All models

VTP-Large-f16d64

Published Publisher MiniMax

Overview

VTP-Large-f16d64 is a pretrained visual tokenizer and image-feature-extraction model from the VTP research project by MiniMax and Huazhong University of Science and Technology. VTP jointly trains image–text contrastive, self-supervised and image-reconstruction objectives to produce visual representations useful for both understanding and generation. (github.com)

Capabilities and uses

  • Encodes images into visual representations and supports image reconstruction.
  • Can be evaluated for zero-shot and linear-probe image classification, and used as a tokenizer in diffusion-model image-generation workflows. Generation requires a separate generative model; VTP is not a standalone chat or text-to-image service.
  • The official repository provides PyTorch examples and evaluation scripts. The model weights are listed on the official Hugging Face model page. (github.com)

Limitations and license

The project is described as work in progress. The repository reports research benchmarks for the VTP framework, which should not automatically be treated as guarantees for every deployment or configuration. The model card identifies a modified MIT license: qualifying commercial products or services exceeding the stated user or revenue thresholds must prominently display “MiniMax VTP”; consult the official license for the full terms. (github.com)

Release date

December 16, 2025. The official announcement states that the technical report and pretrained weights were released on that date. (github.com)

Pricing

As of November 17, 2026, no official input or output API price for this specific model was verified in either MiniMax's mainland China CNY pricing documentation or its international USD pricing documentation. No billing unit or amount is established for it; prices for other MiniMax models do not apply by implication. Self-hosting may incur separate infrastructure costs. (platform.minimax.cn)

Official sources