FastH3
FastH3 (Omni, references to video) needs 25.6 GB of VRAM in ComfyUI, at the diffusion stage of its own template, by file sizes read 3 November 2026. The smallest card that loads it with room to work is the RTX 5090 (32 GB), and the template downloads 41.3 GB.
- Made by
- FastVideo (Hao AI Lab)
- Licence
- MiniMax H3 Community License
- Released
- October 2026
- Sizes read
- 3 November 2026
The licence does not apply in the United States, the European Union, the United Kingdom and South Korea. If you are there, this licence does not cover you.
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Which file on your card
For each memory size, the best file for that card: the highest-precision diffusion file that loads with room to work (one that only just fits when nothing else does), and what is left for working memory while it samples. On a larger card that can be above the file ComfyUI's template loads, which the answer above is about. "Streams" means even the smallest file is larger than the card: ComfyUI still runs it from system RAM, more slowly. FP4 files go only to the Blackwell cards that run them natively.
Omni, references to video
ComfyUI's template for Omni, references to video: INT8 ConvRot, pruned diffusion model, NVFP4 AWQ text encoder. Stages: text encoder 15.7 GB, diffusion 25.6 GB. Download 41.3 GB.
Computed 3 November 2026V2, text or frames to video
ComfyUI's template for V2, text or frames to video: INT8 ConvRot, pruned diffusion model, NVFP4 AWQ text encoder. Stages: text encoder 15.7 GB, diffusion 25.5 GB. Download 41.2 GB.
Computed 3 November 2026Check another card, or pick a file by hand, in the ComfyUI VRAM checker. Built your own workflow around it, with LoRAs or custom nodes? Drop it into the ComfyUI workflow checker.
Where to run it
Omni, references to video
On-demand prices before tax, read 19 October 2026.
V2, text or frames to video
On-demand prices before tax, read 19 October 2026.
About the model
FastH3 is FastVideo's open video model with sound. FastVideo's eight-step distillations of MiniMax H3, from Hao AI Lab at UC San Diego: V2 (September 2026) for text and first-and-last-frame video, and Omni (October 2026), which takes up to nine images, three videos and three audio clips as references. Both make synchronised audio with the picture and use H3's text encoder and VAEs.
Omni, references to video: 33B · V2, text or frames to video: 33B Text encoder: Qwen3-VL 32B MiniMax H3 Community License Released October 2026
Before you download
Where the licence applies, commercial use is allowed; products above US$20 million a year need MiniMax's written authorisation, and commercial interfaces must show "MiniMax H3". The licence does not apply at all in the United States, the European Union, the United Kingdom and South Korea. Read the MiniMax H3 Community License itself before using it for work.
Good at
- Eight sampling steps, against the 20 of ComfyUI's MiniMax H3 templates.
- References in one model: Omni takes up to nine images, three videos and three audio clips.
- An NVFP4 build of Omni from FastVideo for Blackwell cards, about a third the size of BF16.
- H3's other files are shared: with MiniMax H3 set up, the distilled transformer is the only new download.
Watch out for
- MiniMax H3's licence applies, and it excludes the United States, the European Union, the United Kingdom and South Korea; FastVideo's cards do not say so.
- ComfyUI's guide says FastH3 trades some motion and audio fidelity for speed.
- The step count is fixed at eight, and V2 and Omni use different video shifts, 10 and 12: V2's recipe is not Omni's.
- Omni's template needs ComfyUI 0.43.0 or later and has no ComfyUI guide yet; the guide covers V2.
- The NVFP4 build needs a Blackwell card, and no ComfyUI template loads it yet.
Every FastH3 file ComfyUI loads
Sizes are the exact byte counts Hugging Face reports, read 3 November 2026. Files marked "template" are the ones ComfyUI's official workflow downloads.
| Part | Precision | Size | Version | File |
|---|---|---|---|---|
| Diffusion model | BF16, pruned | 44.24 GB | Omni, references to video | fasth3_omniref_8step_pruned_bf16.safetensors |
| Diffusion model | INT8 ConvRot, pruned · template | 22.21 GB | Omni, references to video | fasth3_omniref_8step_pruned_int8_convrot.safetensors |
| Diffusion model | NVFP4 | 13.77 GB | Omni, references to video | fasth3_omniref_qad_nvfp4_comfy.safetensors |
| Diffusion model | BF16, pruned | 44.08 GB | V2, text or frames to video | fastvideo_fasth3_8step_v2_pruned_bf16.safetensors |
| Diffusion model | INT8 ConvRot, pruned · template | 22.13 GB | V2, text or frames to video | fastvideo_fasth3_8step_v2_pruned_int8_convrot.safetensors |
| Text encoder | BF16 | 51.51 GB | all | qwen3vl_32b_minimax_h3_bf16.safetensors |
| Text encoder | INT8 ConvRot | 27.14 GB | all | qwen3vl_32b_minimax_h3_int8_convrot.safetensors |
| Text encoder | NVFP4 AWQ · template | 15.69 GB | all | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
| VAE | INT8 ConvRot · template | 2.81 GB | all | minimax_h3_video_vae_int8_convrot.safetensors |
| Audio VAE | FP32 · template | 0.61 GB | all | minimax_h3_audio_vae_fp32.safetensors |
What FastVideo says it needs
- "Budget about 69 GiB for the DiT and encoder when both stay resident, before the decoder and activations." (FastVideo's own runtime, not ComfyUI, with nothing offloaded, source)
- "Layerwise offload is mandatory: the 47 GB FP8 DiT cannot be resident on 24 GB." (FastVideo's own runtime on an RTX 4090, with its FP8 weights, source)
Those figures and this page measure different things. A vendor's figure covers its own script at its default resolution and length, working memory included. This page counts the weights each stage holds, which is exact, and leaves the working memory as the space that remains, because it depends on your resolution, frame count and attention kernel and nobody publishes it per model. If generation runs out of memory with a file that loads, lower the resolution or the frame count before stepping down a precision. ComfyUI's README says it "can run even the biggest open source models on as low as 4GB vram + 8GB ram" by streaming weights (source), so "streams" means slower, not impossible.
LoRA training for FastH3
No trainer we checked publishes a VRAM figure for training FastH3 yet (kohya's sd-scripts and musubi-tuner, ostris' ai-toolkit, diffusion-pipe, and FastVideo's own repo). When one does, it goes here.
Questions
- How much VRAM does FastH3 need?
- In ComfyUI's own workflow for Omni, references to video, the largest stage is the diffusion stage at 25.6 GB, and the whole workflow downloads 41.3 GB. Working memory for generation comes on top and grows with resolution and length. FastVideo itself says: "Budget about 69 GiB for the DiT and encoder when both stay resident, before the decoder and activations." (FastVideo's own runtime, not ComfyUI, with nothing offloaded).
- Can I run FastH3 on 8 GB of VRAM?
- Not fully. The smallest file here, INT8 ConvRot, pruned at 22.2 GB, needs 25.6 GB at its largest stage, more than 8 GB holds. ComfyUI will still run it by streaming weights from system RAM, which works but is slower.
- Can I run FastH3 on 12 GB of VRAM?
- Not fully. The smallest file here, INT8 ConvRot, pruned at 22.2 GB, needs 25.6 GB at its largest stage, more than 12 GB holds. ComfyUI will still run it by streaming weights from system RAM, which works but is slower.
- Can I run FastH3 on 16 GB of VRAM?
- Not fully. The smallest file here, INT8 ConvRot, pruned at 22.2 GB, needs 25.6 GB at its largest stage, more than 16 GB holds. ComfyUI will still run it by streaming weights from system RAM, which works but is slower.
- Which FastH3 file should I download?
- The highest precision that loads fully on your card. On a 24 GB card that is none of them: even INT8 ConvRot, pruned streams. ComfyUI's template downloads INT8 ConvRot, pruned with the NVFP4 AWQ text encoder. FP4 files (NVFP4) are made for Blackwell cards (RTX 50, RTX PRO 6000, DGX Spark); on older cards, use FP8 or INT8.
- FastH3 or MiniMax H3?
- FastH3 is MiniMax H3 distilled to eight sampling steps, with the same text encoder, VAEs and licence; its transformer file is slightly larger, because it adds gate layers. It finishes in far fewer steps, and ComfyUI's guide says it gives up some motion and audio fidelity for that. For the best quality, MiniMax H3; for fast drafts, FastH3.
- Files and sizes: FastVideo/FastH3-Omni-Comfy on Hugging Face, plus the vendor's own ComfyUI-ready files, read from the Hugging Face API. Template files: ComfyUI's FastH3 guide.
- Model and licence: FastVideo/FastVideo-FastH3-Omni-8-Step-V1 and the MiniMax H3 Community License.
- Card memory: manufacturers' specifications, as on each GPU page. The checker covers NVIDIA cards; ComfyUI also runs on AMD and Apple silicon, where file formats and speed differ and we have not sized them.
- Fit rule: a stage loads fully under 85% of usable memory and is tight up to 95%, the same margins the LLM pages use.
- Prices: each provider's own pricing page, on the dates beside them.