MiniMax H3 VRAM requirements: GPU and system RAM guide
MiniMax H3 local guide: VRAM per tier, GGUF Q4_K_M/Q5_K_M sizes, system RAM, AMD support, generation speed, and when cloud beats local.
On this page
MiniMax H3 VRAM requirements: GPU and system RAM guide
MiniMax H3 is an open-weights omni-modal video model from MiniMax. It generates video with native stereo audio at up to 2K resolution and up to 15 seconds of duration. The released open checkpoints are H3-Base-FL2VA for text-to-video and image-to-video workflows, and H3-Base-Ref2VA for reference-guided generation MiniMax open source announcement. The base checkpoint outputs at 768p; the full 2K path uses H3-Regenerate-2K, which is not released as a local checkpoint MiniMax H3 technical model card.
This guide answers the practical questions: what VRAM do you need, what fits in 6/8/12/16/24/32 GB, which quantized builds work, how much system RAM is required, whether AMD cards are supported, how fast generation is per tier, and when cloud options beat local.
TL;DR
- 16 GB VRAM + 64 GB system RAM is the practical starting point for local H3 with quantized weights and CPU offloading.
- 8–12 GB VRAM is experimental only: GGUF Q3_K_M/Q4_K_M, low resolutions, short clips, heavy offloading, slow generation.
- 24–32 GB VRAM is the comfortable tier; 32 GB is not overkill for Ref2VA or less offloading friction.
- AMD consumer Radeon is not verified for local H3; SGLang validation covers Instinct MI300X/MI355X only.
- RTX 3060 (12 GB) runs quantized H3 under a hard VRAM cap, about 10.6 seconds per denoise step, but not a comfortable daily driver.
- Cloud (Civitai generator, MiniMax API) beats local for 2K output, fast iteration, or when you do not want to manage 42.5–63.5 GB of model files.
Can MiniMax H3 run locally?
Yes. MiniMax H3 can run locally because MiniMax released H3-Base checkpoints as open weights on Hugging Face MiniMax open source announcement. The downloadable checkpoints are BF16 precision and include FL2VA (text-to-video and first/last-frame video) and Ref2VA (reference-guided video) MiniMax H3 technical model card.
But "runs locally" does not mean "runs comfortably on any consumer GPU." The H3 Omni Transformer has 33B parameters MiniMax H3 technical model card, and running it means fitting weights, activations, KV cache, and VAE decode memory in VRAM, with system RAM holding the rest.
There is no single official minimum VRAM figure from MiniMax MiniMax H3 requirements guide. Practical feasibility depends on weight format (BF16, INT8, GGUF), workflow (T2V vs Ref2VA), resolution, clip length, and CPU offloading.
VRAM requirements by tier
The table below summarizes practical VRAM fit for each consumer tier. These are practical estimates based on published deployment results and community reports, not official MiniMax minimums.
| VRAM tier | Fit level | What runs | Required compromises | Source |
|---|---|---|---|---|
| 6 GB | Not practical | Nothing usable | Cannot load even the smallest GGUF text encoder | MiniMax H3 requirements guide |
| 8 GB | Experimental | GGUF Q3_K_M only | Very low resolution, 4–5 sec clips, aggressive CPU offloading, slow generation | MiniMax H3 requirements guide |
| 12 GB | Experimental | GGUF Q3_K_M/Q4_K_M | Hard VRAM cap, ~10.6 s/step denoise, CPU offloading required | SGLang MiniMax-H3 deployment guide |
| 16 GB | Workable | INT8, GGUF Q4_K_M/Q5_K_M | 64 GB system RAM recommended, CPU offloading, preview resolution first | MiniMax H3 requirements guide |
| 24 GB | Practical | INT8, GGUF Q5_K_M/Q6_K | Better headroom for longer clips and higher resolution | MiniMax H3 requirements guide |
| 32 GB+ | Recommended | INT8, Ref2VA, BF16 possible | Less offloading friction, better for Ref2VA workflows | MiniMax H3 requirements guide |
8 GB VRAM
Can MiniMax H3 run on 8GB VRAM? Only experimentally. The smallest GGUF variant (Q3_K_M at 8.9 GB for FL2VA) cannot fit entirely in 8 GB VRAM, so heavy CPU offloading is required MiniMax H3 Pruned GGUF README. Expect very low resolution, short clips, and generation measured in minutes per clip rather than seconds. ComfyUI does not release the text encoder from host RAM, which adds pressure on system RAM HF Comfy-Org/MiniMax-H3 discussion #6.
12 GB VRAM (RTX 3060, RTX 4070, RTX 5070)
Can RTX 3060 run MiniMax H3? Yes, but only with quantized weights under a hard VRAM cap. SGLang measured an RTX 3060-class 12 GB card at ~10.6 seconds per denoise step with ~9.4 seconds decode, roughly 235 seconds per request at 480P, ahead of ComfyUI under the same cap SGLang MiniMax-H3 deployment guide. Ampere cards (RTX 3060, RTX 3090) land above the 40-series step times measured in that guide. This is a functional but slow path; it is not a comfortable daily-driver card for video generation.
16 GB VRAM
16 GB is the practical starting point for compact INT8 or GGUF Q4_K_M workflows, preferably with at least 64 GB of system RAM MiniMax H3 requirements guide. SGLang measured a 16 GB consumer card at ~11.9 seconds per denoise step, ~11 seconds decode, roughly 250 seconds per request under a hard 16 GiB cap SGLang MiniMax-H3 deployment guide.
24 GB VRAM (RTX 4090, RTX 3090)
24 GB is better suited to local generation, longer clips, and higher working resolutions while still relying on memory management MiniMax H3 requirements guide. SGLang measured a 24 GB card at ~8.5 seconds per denoise step with six resident layers, roughly 249–260 seconds per request in ComfyUI comparison SGLang MiniMax-H3 deployment guide.
32 GB VRAM (RTX 5090)
Is 32 GB of VRAM overkill? Not if you use Ref2VA or want fewer offloading compromises. SGLang measured an RTX 5090 at 5.14–5.17 seconds per step with the DiT pinned, roughly 112 seconds per request at 864×480 / 124 frames / 20 steps SGLang MiniMax-H3 deployment guide. 32 GB gives headroom for Ref2VA's larger working set and reduces drive traffic per step.
Quantized builds and file sizes
Storage size is separate from runtime VRAM. A 21 GB model file does not mean 21 GB of VRAM is sufficient MiniMax H3 requirements guide.
Official BF16 checkpoints
| Component | File | Size | Destination |
|---|---|---|---|
| FL2VA diffusion | minimax_h3_fl2va_pruned_int8_convrot.safetensors | 21 GB | models/diffusion_models/ |
| Text encoder | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | 15.7 GB | models/text_encoders/ |
| Video VAE | minimax_h3_video_vae_fp16.safetensors | 5.21 GB | models/vae/ |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors | 0.61 GB | models/vae/ |
| Ref2VA diffusion | minimax_h3_ref2va_pruned_int8_convrot.safetensors | 21 GB | models/diffusion_models/ |
T2V/I2V total: 42.5 GB. With Ref2VA: 63.5 GB MiniMax H3 requirements guide.
Community GGUF quantized builds
Community GGUF builds compress the pruned weights to smaller files. The Abiray pruned GGUF repository lists these variants MiniMax H3 Pruned GGUF README:
| Quantization | FL2VA size | Ref2VA size | Target hardware |
|---|---|---|---|
| Q3_K_M | 8.9 GB | 8.9 GB | Low VRAM (~12 GB GPUs) |
| Q4_K_M | 11.6 GB | 11.6 GB | Recommended balance (16 GB GPUs) |
| Q4_K_S | 11.6 GB | 11.6 GB | Compact 4-bit K-Quant |
| Q5_K_M | 14.1 GB | 14.1 GB | High quality balance (24 GB GPUs) |
| Q5_K_S | 14.1 GB | 14.1 GB | Compact 5-bit K-Quant |
| Q6_K | 16.7 GB | 16.7 GB | Near-lossless precision |
| Q8_0 | 21.6 GB | 21.6 GB | Maximum fidelity |
The GGUF README recommends Q4_K_M for 16 GB GPUs and Q5_K_M for 24 GB GPUs MiniMax H3 Pruned GGUF README.
INT8 and NVFP4
The official FL2VA diffusion weights ship as pruned INT8 convrot safetensors (21 GB each). The text encoder ships as Qwen3-VL-32B with NVFP4 AWQ quantization (15.7 GB) MiniMax H3 requirements guide. Community or third-party FP8, INT8, and GGUF quantizations exist but are not MiniMax's official release MiniMax H3 technical model card.
System RAM requirements
System memory becomes increasingly important as model components move out of VRAM. ComfyUI never releases the text encoder from host RAM HF Comfy-Org/MiniMax-H3 discussion #6.
| System RAM | Fit level | Notes |
|---|---|---|
| Less than 32 GB | Not recommended | Insufficient for CPU offloading |
| 32 GB | Experimental | Minimal headroom |
| 64 GB | Workable | Practical compact-workflow starting point |
| 96 GB | Practical | More offloading headroom |
| 128 GB+ | Recommended | Best for Ref2VA and large outputs |
64 GB is a more practical compact-workflow starting point than 32 GB MiniMax H3 requirements guide. One ComfyUI reporter found a --fast-disk flag reduced anonymous RSS from 45.4 GiB to 12.6 GiB on a 16 GB VRAM card, moving weights into reclaimable page cache HF Comfy-Org/MiniMax-H3 discussion #6.
AMD support status
Does MiniMax H3 work on AMD? SGLang validation covers AMD Instinct MI300X and MI355X SGLang MiniMax-H3 deployment guide. That is data-center Instinct hardware, not consumer Radeon cards. The MiniMax H3 requirements guide explicitly states AMD Radeon is not verified for local MiniMax H3 MiniMax H3 requirements guide. Do not interpret Instinct validation as confirmed support for consumer Radeon GPUs.
Realistic generation-speed expectations
Generation speed depends on VRAM tier, offloading aggressiveness, resolution, clip length, and storage speed. All published figures below are anchored at 480P; activations grow with pixel count, so 768P requires dropping resident DiT layers to zero first SGLang MiniMax-H3 deployment guide.
| VRAM | Denoise step | Decode | Approx. per request | Notes | Source |
|---|---|---|---|---|---|
| 12 GB | ~10.6 s | ~9.4 s | ~235 s | Hard 12 GiB cap, 480P | SGLang MiniMax-H3 deployment guide |
| 16 GB | ~11.9 s | ~11 s | ~250 s | Hard 16 GiB cap, 480P | SGLang MiniMax-H3 deployment guide |
| 24 GB | ~8.5 s | ~9.6 s | ~249–260 s (ComfyUI) | Six resident layers, 480P | SGLang MiniMax-H3 deployment guide |
| 32 GB | 5.14–5.17 s | ~6.4 s | ~112 s | DiT pinned, 864×480, 124 frames, 20 steps | SGLang MiniMax-H3 deployment guide |
Community reports on the HF discussion thread describe 4-second 720p videos taking 15:01 on first run, then 416–422 seconds on subsequent runs with quantized versions HF Comfy-Org/MiniMax-H3 discussion #6. Another user reported 10–13 minutes for 12-second 16×9 HD videos on an RTX 4090 with 128 GB RAM HF Comfy-Org/MiniMax-H3 discussion #6.
Storage speed matters. A PCIe 3.0 drive (~3.2 GiB/s) can become the denoise clock; a SATA SSD is not usable for this model SGLang MiniMax-H3 deployment guide.
ComfyUI workflow support
ComfyUI 0.30+ has native MiniMax-H3 workflows MiniMax H3 requirements guide. The GGUF path requires ComfyUI-GGUF custom nodes and the UnetLoaderGGUF node MiniMax H3 Pruned GGUF README. Internal links to related planned site content: MiniMax H3 ComfyUI setup, MiniMax H3 local deployment, MiniMax H3 API guide.
When cloud beats local
Cloud options (Civitai generator, MiniMax API) beat local when:
- You want 2K output: the downloadable H3-Base checkpoint generates 768p; full 2K requires H3-Regenerate-2K, which is not a local checkpoint MiniMax H3 technical model card.
- You need fast iteration: local generation runs in minutes per clip; hosted generators return faster.
- You do not want to manage 42.5–63.5 GB of model files and 64+ GB of system RAM.
- Your GPU is below 16 GB VRAM or is an unsupported AMD consumer card.
Hosted access follows a separate policy from open-weight downloads MiniMax open source announcement.
Common errors and fixes
| Error | Cause | Fix | Source |
|---|---|---|---|
| OOM on model load with 16 GB VRAM | Loading BF16 checkpoint (42.5 GB) without quantization | Use GGUF Q4_K_M or INT8 pruned convrot files instead | MiniMax H3 requirements guide, MiniMax H3 Pruned GGUF README |
| Slow generation (~10 min per clip) on 12 GB VRAM | Hard VRAM cap causing frequent disk reads | Use --fast-disk flag in ComfyUI to move weights into reclaimable page cache; ensure 64 GB+ system RAM HF Comfy-Org/MiniMax-H3 discussion #6 | |
| Text encoder not released from host RAM | ComfyUI retains text encoder in host memory | Use --cache-none flag only for node-output caching; use --fast-disk for weight residency HF Comfy-Org/MiniMax-H3 discussion #6 | |
| 2K output not available locally | H3-Base generates 768p; 2K requires H3-Regenerate-2K | Accept that 2K is hosted-only; use Civitai generator or MiniMax API for 2K MiniMax H3 technical model card | |
| AMD Radeon card not working | SGLang validation covers AMD Instinct MI300X/MI355X only, not consumer Radeon | Use NVIDIA CUDA consumer GPU for local MiniMax H3; AMD consumer cards are unsupported MiniMax H3 requirements guide, SGLang MiniMax-H3 deployment guide | |
| Generation stalls on SATA SSD | Storage speed below PCIe 3.0 threshold | Use at least a PCIe 3.0 NVMe drive; SATA SSD is not usable for this model SGLang MiniMax-H3 deployment guide | |
| Quality drops with Q3_K_M GGUF | Lowest quantization drops detail | Use Q4_K_M for balanced quality/size; Q5_K_M for high quality on 24 GB GPUs MiniMax H3 Pruned GGUF README | |
| VRAM insufficient even with Q4_K_M | System RAM too low for CPU offloading | Ensure at least 64 GB system RAM; --fast-disk helps on lower-RAM machines HF Comfy-Org/MiniMax-H3 discussion #6 |
FAQ
Can MiniMax H3 run on 8GB VRAM? Only experimentally. The smallest GGUF variant (Q3_K_M at 8.9 GB) cannot fit entirely in 8 GB VRAM, so heavy CPU offloading is required. Expect very low resolution, short clips, and slow generation MiniMax H3 Pruned GGUF README, MiniMax H3 requirements guide.
Can RTX 3060 run MiniMax H3? Yes, but only with quantized weights under a hard VRAM cap. SGLang measured an RTX 3060-class 12 GB card at roughly 10.6 seconds per denoise step, about 235 seconds per request at 480P. It is functional but slow SGLang MiniMax-H3 deployment guide.
What are the minimum GPU requirements to run MiniMax H3? MiniMax does not publish an official minimum VRAM figure. Practically, 16 GB VRAM with 64 GB system RAM is the workable starting point using quantized weights and CPU offloading. 8–12 GB VRAM is experimental only MiniMax H3 requirements guide.
Does MiniMax H3 work on AMD? SGLang validation covers AMD Instinct MI300X and MI355X, which are data-center cards. Consumer Radeon cards are not verified for local MiniMax H3 SGLang MiniMax-H3 deployment guide, MiniMax H3 requirements guide.
Is 32 GB of VRAM overkill? Not if you use Ref2VA or want fewer offloading compromises. SGLang measured an RTX 5090 (32 GB) at 5.14–5.17 seconds per step with the DiT pinned. For standard FL2VA workflows, 24 GB is already practical SGLang MiniMax-H3 deployment guide, MiniMax H3 requirements guide.
Can you run MiniMax H3 locally? Yes. MiniMax released H3-Base checkpoints as open weights. The base checkpoint generates 768p video with native stereo audio at up to 15 seconds. Full 2K output requires H3-Regenerate-2K, which is not released as a local checkpoint MiniMax open source announcement, MiniMax H3 technical model card.
How much system RAM do you need? Less than 32 GB is not recommended. 64 GB is the practical compact-workflow starting point. 96–128 GB gives CPU offloading more headroom for Ref2VA and large outputs MiniMax H3 requirements guide.
What GGUF quantization should I use? Q4_K_M is recommended for 16 GB GPUs. Q5_K_M is recommended for 24 GB GPUs. Q3_K_M targets ~12 GB GPUs but trades quality for size MiniMax H3 Pruned GGUF README.
How do I add MiniMax H3 to ComfyUI? ComfyUI 0.30+ ships native MiniMax-H3 workflows. The GGUF path additionally needs the ComfyUI-GGUF custom nodes and the UnetLoaderGGUF node MiniMax H3 requirements guide, MiniMax H3 Pruned GGUF README.
Is MiniMax H3 Max open source? H3-Base is open weights. H3 Max is the hosted post-trained model MiniMax released with fal.ai, and hosted access follows a separate policy from open-weight downloads MiniMax open source announcement.
How long can MiniMax H3 videos be? Up to 15 seconds, integer values only. The open H3-Base checkpoint generates 768p with native stereo audio; full 2K requires H3-Regenerate-2K, which is hosted MiniMax open source announcement, MiniMax H3 technical model card.
Sources
- MiniMax H3 Requirements: VRAM, RAM, GPU & Storage — Practical VRAM and RAM tier guidance, model file sizes, ComfyUI workflow support, AMD status.
- MiniMax H3 Technical Model Card — Architecture (33B Omni Transformer, H3-Base, H3-Regenerate-2K), BF16 checkpoints, 768p base / 2K hosted, FL2VA vs Ref2VA.
- MiniMax H3 Open Source Announcement — Open weights release, license, output specifications, workflow recommendations.
- SGLang MiniMax-H3 Deployment Guide — Verified per-request timings by VRAM tier, storage speed requirements, AMD Instinct validation.
- MiniMax H3 Pruned GGUF README — GGUF quantization sizes (Q3_K_M through Q8_0), target hardware, ComfyUI setup.
- HF Comfy-Org/MiniMax-H3 Discussion #6 — Community VRAM/RAM reports, generation speed anecdotes,
--fast-diskflag behavior.