Model reference · open weights

VisionReward-Video

LLMs zai-org Text gen 1 build Its own licence terms 3k dl/mo

VisionReward-Video is an open-weight language model from zai-org. VisionReward-Video (BF16) weighs 25.0 GB; the smallest configuration that runs it is 2× RTX 4060 Ti 16 GB.

Request Rent a GPU

For self-hosting, compare this build's requirements below with the GPU you plan to rent.

VisionReward-Video is a 12.5B parameter text-generation model developed by zai-org for aligning visual generation models with human preferences.
It supports a context length of 2048 tokens and operates in English under an other licence.
The model uses a multi-dimensional framework to decompose human preferences into judgment questions for scoring video quality.

Summary of the zai-org/VisionReward-Video model card, 2026-10-01

Hugging Face ↗GitHub ↗

What it is

Released by	zai-org
Released	2024-12-10
Parameters	12.5B
VRAM	25.0 GB for the weights

What it runs on

Memory and cards for VisionReward-Video (BF16)

25.0 GBweights, file size

524 MBcache per 1K tokens

2.0 GBruntime overhead, at least

2,048 tokenscontext max

Card	Requests at once	Context max	Memory
Card	2K each	Context max	Memory
RTX 3060 12 GB … RTX 4090 24 GB 4 smaller cards
RTX 5090 32 GB	3	all 2K	31.0 GB
L40S 48 GB	15	all 2K	44.0 GB
A100 80 GB	47	all 2K	78.2 GB
H100 80 GB	43	all 2K	78.1 GB
RTX PRO 6000 Blackwell 96 GB	58	all 2K	93.8 GB
DGX Spark (GB10) 128 GB unified	70	all 2K	107 GB
H200 141 GB	99	all 2K	138 GB
B200 180 GB	134	all 2K	176 GB
2× RTX 4060 Ti 16 GB tensor parallel	1	all 2K	15.4 GB a card
2× RTX 4090 24 GB tensor parallel	16	all 2K	23.4 GB a card
2× RTX 3090 24 GB tensor parallel	16	all 2K	23.4 GB a card
2× RTX 5090 32 GB tensor parallel	30	all 2K	31.0 GB a card

Memory needed at each load

Requests at once	2K tokens each
1	28.1 GB
5	32.4 GB
8	35.6 GB
16	44.2 GB
32	61.4 GB
64	95.8 GB

One card, with vLLM's small-card settings.

From the model card

What zai-org says about VisionReward-Video

Read the model card

Introduction

We present VisionReward, a general strategy to aligning visual generation models——both image and video generation——with human preferences through a fine-grainedand multi-dimensional framework. We decompose human preferences in images and videos into multiple dimensions,each represented by a series of judgment questions, linearly weighted and summed to an interpretable and accuratescore. To address the challenges of video quality assess-ment, we systematically analyze various dynamic features of videos, which helps VisionReward surpass VideoScore by 17.2% and achieve top performance for video preference prediction. Here, we present the model of VisionReward-Video.

Using this model

You can quickly install the Python package dependencies and run model inference in our github.

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

Details

Type	Language models
Task	Text gen
Context	2,048 tokens
Popularity	3k downloads / month
Weights	25.0 GB, VisionReward-Video (BF16, file size)

Languages