Skip to content
Video

H3 Turbo Runtime

Low-VRAM Inference Engine for MiniMax H3

3 min read

This is not a newly trained model, but an optimized inference engine built by developer quifox for the MiniMax-H3 video-and-audio generation model. The solution combines community W4A8 quantized weights, the lightx2v/Minimax-h3-Turbo acceleration adapter (which reduces generation steps), and custom Triton compute kernels. Compared to the unmodified baseline PyTorch setup, the engine radically cuts video memory usage and speeds up rendering. The tool is useful for anyone wanting to generate videos with synchronous audio locally on consumer GPUs with 6–8 GB of VRAM, as well as within ComfyUI and CozyUI workflows.

Key features

  • Two memory modes: the Stream strategy uses weight offloading to fit in about 4.0–4.2 GB of VRAM, while Fast uses custom Triton kernels for maximum speed.
  • Step-by-step component offloading: the Qwen3-VL INT8 text encoder is fully unloaded from memory before the main diffusion model is loaded, and video and audio decoding runs sequentially.
  • Flexible formats and duration: resolutions of 832×480 (16:9), 960×544 (16:9), 480×832 (9:16), 624×624 (1:1), and 720×544 (4:3) are supported, along with generating clips up to 15 seconds in a single pass at 24 fps.
  • Automatic dependency downloading: the repository itself does not store weights, but includes a script to download all 6 external components with SHA-256 hash verification and resume support.

What you can do

  • Generate video with audio on a laptop: render 5-second clips at 832×480 or 960×544 resolution on RTX 4060 Laptop 8GB class GPUs in 2.5–3.5 minutes.
  • Create content for different platforms: assemble 9:16 vertical clips, 1:1 square posts, or long 15-second scenes in a single pass.
  • Character-driven generation (Ref2VA): replace the base weights and LoRA with the bundle from the Kutches/minmax repository and the 768p Ref2V LoRA to work with multiple character references.

Example

uv run python deploy.py render \
  --comfy /path/to/ComfyUI \
  --models weights/ \
  --condition path/to/condition.safetensors \
  --duration 5 --width 832 --height 480 \
  --steps 4 --strategy fast --out runs/demo
Read moreDetails
Benchmarks (5 sec, 832×480): on a Tesla T4 16GB, peak usage dropped from 14,649 MiB (2,287.9 s) to 4,043 MiB (227.18 s for 4 steps in Stream mode); on an RTX 4060 Laptop 8GB, rendering takes 161.54 s at a peak of 5,695 MiB.
Generating a 15-second clip (360 frames) on an RTX 4060 Laptop 8GB takes 694.7 s at 6,995 MiB VRAM.
Requirements: an NVIDIA GPU (Turing, Ampere, Ada, and data center) with 6 GB VRAM or more and ComfyUI installed.
Licenses: the engine code is distributed under GPL-3.0-only, and external weights fall under the MiniMax H3 Community License Agreement.
Discuss Telegram 0

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # quifox/H3-Turbo-Runtime
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download quifox/H3-Turbo-Runtime --local-dir ~/models/H3-Turbo-Runtime
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to H3-Turbo-Runtime
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • SageAttention 3v2.2.0from October 28, 2025

    even faster on RTX 50 / Blackwell

    FP4 attention for Blackwell. Faster than SageAttention 2, but the authors suggest v2 when accuracy matters.

    How to enable: Build from the sageattention3_blackwell folder of the repo; in ComfyUI via KJNodes.

    Earlier generations (2)
    • SageAttention 2++ — Quantised (INT8/FP8) attention with no visible quality loss. The main speed-up for image and video generation.
    • SageAttention 1 — First version: INT8 for Q·K. Still works, but 2++ is faster and more accurate.
  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (3)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
    • xFormers — The older way to speed up attention, works almost everywhere. Sage or Flash are faster where available.
  • cache-dit (DBCache)v1.5.6from November 20, 2026new release

    1.5–3× at nearly the same quality

    Reuses block outputs between diffusion steps. Supports most DiT models.

    How to enable: pip install cache-dit; one line cache_dit.enable_cache(pipe) for diffusers.

    Earlier generations (2)
    • MagCache — Skips steps by magnitude of change. More accurate than TeaCache, no calibration.
    • TeaCache — The first popular diffusion step cache. Works, newer methods are more accurate.
  • LightX2V step distill0.6.1from November 9, 2026new release

    5–10× fewer steps

    Distilled LoRAs and an engine: video in 4–8 steps instead of 30–50.

    How to enable: Load their step-distill LoRA, CFG 1.0, 4–8 steps.

  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

MiniMax H3 prompts

All prompts

Ready prompts with the resulting video: copy and try.

Model family

New posts about versions of H3-Turbo-Runtime (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

quifox/H3-Turbo-Runtime
Author
quifox
Category
Video
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments