Running MiniMax H3 on 96GB RTX PRO 6000 via Gemini Pro

The article documents a hands-on experiment using a Gemini Pro subscription ($20/month) to access Google Colab's high-end GPU offering: an RTX PRO 6000 Blackwell with 96GB VRAM (G4 instance). The author previously showed that Pro users can access A100 and G4 GPUs, while Ultra users get H100. The G4's 96GB VRAM makes it suitable for large video generation models.

MiniMax H3 Architecture

MiniMax H3 jointly generates video and audio from a single backbone. The core generation backbone, H3-Omni-Transformer, has 33B parameters and uses a dense single-stream architecture where text, video, and audio features are predicted jointly in one sequence, then decoded by separate VAEs. The text encoder uses Qwen3-VL-32B pretrained weights (layer 50 representations as conditioning). The video VAE compresses spatially by 16× and temporally by 4×. About 13B parameters belong to AdaLN modulation branches that dynamically adjust internal features; during inference these modulation results can be precomputed and cached in system memory, not VRAM. The ComfyUI community provides a pruned version based on this optimization, distinct from full checkpoints used for continued training.

Model Size and Official Deployment Paths

According to official metadata, the raw BF16 checkpoints total 144.02GB: backbone 66.28GB, encoder 66.71GB, plus two VAE sets. The full system comprises three pipelines: Context-IR for input preparation, Base for 768p generation, and Regenerate-2K for super-resolution. Only the Base branch is open-sourced. Official documentation lists deployment options:

Base original BF16 : Diffusers places encoder and backbone on two 80GB GPUs; single 80GB GPU requires CPU offload.

Consumer GPU deployment : 24–32GB GPUs with INT8 quantization of backbone and encoder, streaming weights from CPU memory; requires large system RAM.

This experiment (quantized Base) : Single 96GB G4 GPU with ~176GiB system RAM, generating 1344×768 15-second video; observed peak VRAM ~54GiB.

CPU offload trades data-movement time for VRAM space; on a large VRAM card like G4 the goal is full-resident execution.

Environment and Deployment

The author uses local Codex with Google's official Colab MCP to connect to the cloud instance. Pro accounts receive ~200 compute units monthly, sufficient for H3 inference tests. The cloud instance comes preinstalled with PyTorch 2.11.0+cu130, matching Comfy-Org's INT8 ConvRot weight requirements. After adding ComfyUI dependencies, the native H3 nodes are invoked via terminal for batch processing without launching the WebUI.

Codex downloads four weight files totaling 40.07GB: backbone and video VAE use pruned INT8 ConvRot, text encoder is quantized, audio VAE stays FP32. Despite the 40GB static weight size, peak VRAM during forward pass reaches 54.08GB. Automated scripts lock commit and revision, verify SHA-256 hashes; download and verification take ~2 minutes 19 seconds.

Experiment Setup

The prompt (designed by Codex) describes a "paper universe" story: a paper boat unfurls a red sail from a lab notebook, flies through ink-drawn nebulae, and lands back on the page. The 15-second video is split into three shots with audio cues: paper rustle, wind whoosh, calm music, ending with a crisp bell, no narration. Fixed seed 20261013, resolution 1344×768, 24 fps (362 frames generated, 15 seconds exported), using res_multistep sampler and simple scheduler.

Benchmark Results

Download & verification : 2 min 19 sec (both Base 20-step and Turbo 8-step)

Sampling time : Base 20-step 18 min 11 sec; Turbo 8-step 7 min 17 sec

Total inference time : Base 20-step 18 min 54 sec; Turbo 8-step 7 min 55 sec

Peak VRAM observed : Base 20-step 54.08 GiB; Turbo 8-step 53.96 GiB

Turbo 8-step reduces sampling time from 18:11 to 7:17, a 2.50× speedup, nearly proportional to step reduction. Including fixed audio/video codec overhead, total time drops 58.1% (saving ~11 minutes). Peak VRAM remains ~54GB; Turbo saves compute time, not resident VRAM.

Qualitative Comparison

Frame-by-frame comparison shows Turbo preserves sharp creases, paper texture, and geometric edges without noticeable smearing; formula and celestial elements appear denser. Base 20-step yields more negative space and softer, more natural light transitions.

The author concludes that for developers needing open-source video generation without a local GPU server, the $20 Gemini Pro subscription providing a 96GB RTX PRO 6000 is a significant advantage.