Local AI Hardware · Video Model Requirements
MiniMax H3 VRAM Calculator: GPU Requirements Checker
Check GPU compatibility before downloading ~498 GB — select your GPU model, GPU count, and ComfyUI diffusion checkpoint to verify whether MiniMax H3 fits fully in VRAM, requires CPU offloading, or is not practical.
By SudoAI Team · Numbers verified against official sources (2026-10-12) · Read the playbook: How to run MiniMax H3 locally
Hardware & Model Selection
Pruned models drop ~13B precomputed AdaLN parameters from the 33B transformer for inference. Text encoder (Qwen3-VL 32B) is loaded and unloaded separately in ComfyUI.
Hardware Verdict & Memory Allocation
How we calculate MiniMax H3 VRAM requirements
MiniMax H3 is a 33B-parameter dense single-stream transformer generating 1344×768 video at 24fps (~15 s) with native stereo audio. Hardware planning relies on verified numbers from the official Diffusers and ComfyUI repositories.
GPU requirement = chosen diffusion model (GiB) + 3 GiB headroom
- Headroom: 3 GiB at 1344×768 resolution is our working allowance for activations and VAE decoding (estimate, not verified).
- Text encoder decoupling: In ComfyUI, the Qwen3-VL 32B text encoder (14.61–47.97 GiB) is loaded to create prompt embeddings and then unloaded from VRAM before diffusion sampling begins.
- Fits fully on GPU: Total available VRAM meets or exceeds the required VRAM (checkpoint + 3 GiB headroom). Model weights remain fully GPU-resident.
- Runs only with offloading: Total VRAM is between 12 GiB and the requirement. Diffusion blocks stream sequentially from system memory. Per official Diffusers guidance, this tier needs ~75 GB host RAM at int8.
- Not practical: Below 12 GiB total VRAM. The official Diffusers minimum tested configuration is 12–16 GB at a reduced canvas (960×544).
- Community measurement: On an RTX 4090 24 GiB, DmitryDB benchmarks observed peak VRAM of 20.42 GiB for int8-convrot-safe and 21.03 GiB for int8-convrot-balanced (community, not official).
- Omitted figures: Per-GPU generation speeds, AMD ROCm execution, and Apple Silicon / MPS compatibility are not verified in official sources and are omitted.
Official reference file sizes & Diffusers memory tiers
Verified figures from official ComfyUI (Comfy-Org/MiniMax-H3) and Diffusers (Diffusers docs) releases. Full Hugging Face repository footprint is ~498 GB.
Official ComfyUI model files (GiB)
| Component | Precision / Variant | Size (GiB) | Role & Notes |
|---|---|---|---|
| Diffusion model | bf16 | 61.73 | Full precision unpruned 33B transformer |
| Diffusion model | int8_convrot | 31.70 | 8-bit quantized unpruned transformer |
| Diffusion model | pruned_bf16 | 37.46 | Drops ~13B precomputed AdaLN parameters |
| Diffusion model | pruned_fp8_scaled | 19.52 | FP8 scaled quantization, pruned architecture |
| Diffusion model | pruned_int8_convrot | 19.53 | Template default; recommended for 24GB GPUs |
| Diffusion model | pruned_w6a8 | 14.89 | Compact 6-bit weight / 8-bit activation quant |
| Text encoder (Qwen3-VL 32B) | bf16 | 47.97 | Full precision text & vision conditioner |
| Text encoder (Qwen3-VL 32B) | int8_convrot | 25.28 | 8-bit quantized text conditioner |
| Text encoder (Qwen3-VL 32B) | nvfp4_awq | 14.61 | Ultra-compact NVFP4 text conditioner |
| Video VAE | fp16 / int8_convrot | 4.85 / 2.62 | Spatial video autoencoder |
| Audio VAE | fp32 | 0.56 | Native stereo audio autoencoder |
| Turbo LoRA | bf16 | 1.82 | Reduces sampling from 20 steps to 8 (T2V/I2V) or 4 (reference) |
Official Diffusers memory guidance tiers
| Hardware Tier | Loading Strategy | System RAM Required | Supported Canvas & Behavior |
|---|---|---|---|
| 2x 80 GB | Full bf16 | Standard | Each large component (61.7 GB transformer, 62.1 GB Qwen3-VL) resident on its own card; native 1344×768. |
| 1x 80 GB | ComponentsManager + CPU offload | Standard | ComponentsManager automatically offloads inactive components between stages. |
| 2x 48 GB | int8 on both large components | Standard | Both transformer and Qwen3-VL loaded in int8 precision across cards. |
| 24–32 GB card | int8 + CPU block streaming | ~75 GB host RAM | int8 transformer + text encoder; streams transformer blocks from CPU RAM. |
| 12–16 GB card | int8 streaming + VAE offload | ~75 GB host RAM | Leaf-level video VAE offload; run on small canvas e.g. 960×544 (~3x faster per step than 1344×768). |
Frequently Asked Questions
What are the MiniMax H3 VRAM and GPU requirements?
MiniMax H3 diffusion weights range from 14.89 GiB (pruned_w6a8) and 19.53 GiB (pruned_int8_convrot) up to 61.73 GiB (unpruned bf16). Running full unpruned bf16 alongside the 62.1 GB Qwen3-VL conditioner requires 2x 80 GB GPUs. For consumer hardware, the recommended baseline is a 24GB GPU (e.g. RTX 3090 / 4090) running the pruned_int8_convrot checkpoint. GPUs with 12–16GB VRAM require CPU block streaming, ~75 GB host RAM, and reduced canvas resolutions such as 960×544.
Can MiniMax H3 run on an RTX 3090 or RTX 4090 24GB GPU?
Yes. The official ComfyUI template default checkpoint (pruned_int8_convrot, 19.53 GiB) fits fully in 24GB VRAM
alongside our estimated 3 GiB headroom at 1344×768. Because ComfyUI loads and unloads the text encoder separately, the conditioner
does not compete for VRAM during diffusion. DmitryDB's independent community testing confirmed an RTX 4090 24 GiB passed with
20.42 GiB peak VRAM in safe mode and 21.03 GiB in balanced mode (community, not official).
Can MiniMax H3 run on 12GB or 16GB GPUs like the RTX 3060 or 4060 Ti?
Yes, but only via offloading. Official Diffusers guidance confirms 12–16 GB GPUs can execute MiniMax H3 using int8 quantization, streaming transformer blocks sequentially from CPU RAM, and applying leaf-level offload to the video VAE. To maintain viable per-step speeds, Diffusers documentation recommends lowering the resolution to 960×544, which runs approximately 3x faster per step than native 1344×768.
How much host RAM (system RAM) does MiniMax H3 need?
Official Diffusers documentation specifies ~75 GB host RAM for int8 offloading workflows. If your GPU has 12–32 GB VRAM and relies on CPU block streaming, having under 75 GB system RAM will likely trigger system out-of-memory errors.
What is the difference between pruned and unpruned checkpoints?
MiniMax H3 features a 33B-parameter dense transformer. Approximately 13B parameters reside in Adaptive LayerNorm (AdaLN) branches that can be precomputed for inference. Pruned checkpoints drop these redundant branches, reducing the bf16 model size from 61.73 GiB down to 37.46 GiB without architectural compromises.
Can I use MiniMax H3 outputs commercially?
No, not under the standard license. The model is distributed under the custom MiniMax H3 Community License Agreement. Commercial use of locally generated outputs requires purchasing a MiniMax commercial license (ComfyUI documentation notes that Comfy is an official reseller).
Primary sources and verified documentation
- MiniMax-H3 on Hugging Face — Official open-weights repository
- MiniMax-H3 GitHub Repository — Official codebase & license agreement
- Diffusers MiniMax H3 Pipeline Documentation — Official memory tiers & CPU offload guidance
- Comfy-Org MiniMax-H3 on Hugging Face — Official ComfyUI quantized files & sizes
- ComfyUI MiniMax H3 Official Tutorial — Templates, sampler configs, and attention notes
- DmitryDB MiniMax-H3 ComfyUI Quants — RTX 4090 24GB peak VRAM measurements (community, not official)
Related playbooks & tools
- How to run MiniMax H3 locally: ComfyUI and Diffusers guide — Step-by-step setup guide
- ChatGPT token counter & OpenAI cost calculator — Client-side API token cost forecasting
- AI coding assistant comparison checklist — Score local and hosted AI tools
- Editorial policy — Strict rules on primary sources and verified evidence