What is HyperFlow H3 Flash?

HyperFlow H3 Flash is a 14B video model built on MiniMax H3. It uses six inference steps and int8 weights to reduce latency while preserving output quality.

How fast is HyperFlow H3 Flash?

It generates 768p video in real time on four NVIDIA B200 GPUs and delivers a 12.7× speedup over the H3 baseline on a single B200.

Can HyperFlow H3 Flash run on consumer GPUs?

Yes. The same checkpoint and six-step settings run on consumer-grade GPUs for local testing and integration, at the speed supported by the hardware.

The third release in the HyperFlow series

HyperFlow H3 Flash

A 14B model that generates video in real time and holds its quality.

Get the code Download weights

14B · 6-Step · int8
768p in Real Time
Runs on Consumer GPUs

Introducing HyperFlow H3 Flash

See it in action

01 Capabilities

Real-time generation

Generates 768p video in real time on four B200 GPUs.

02 Capabilities

Full quality

Built on MiniMax H3, this 14B model uses six inference steps and int8 quantization to accelerate generation while maintaining output quality across different GPU setups.

03 Capabilities

Consumer-grade GPU

The same model checkpoint runs on consumer-grade GPUs for local testing and integration.

How it works

The speed comes from four decisions in how the model is run. On a single B200, HyperFlow H3 Flash delivers a 12.7× speedup over the H3 baseline.

01
Steps
The sampler converges in six passes instead of a longer trajectory. Sampling accounts for most of the wall-clock time of a generation, so reducing the number of passes is the largest single gain.
02
Precision
Weights are quantized to int8, which reduces the memory traffic of each pass. Output quality was verified against the bf16 baseline.
03
Parameters
The 14B parameter count was selected to meet a per-step latency budget while retaining detail at 768p.
04
Hardware
Four B200 GPUs provide the parallelism needed to complete those six passes at the speed of playback. Consumer hardware runs the same six passes at its own pace. The weights and the settings are the same in both cases.

The HyperFlow series

The series now includes three releases, each addressing a different constraint. HyperFlow reduced the cost of training. HyperFlow Infinity extended the length of a single generation. HyperFlow H3 Flash reduces latency until generation and playback run at the same speed.