Pika H3 Flex
Pika H3 Flex (Text or frames to video with audio) needs 26.4 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 42.0 GB.
- Made by
- Pika Labs
- 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; its LoRA adapter counts with the diffusion model. 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; its LoRA adapter counts with the diffusion model. 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.
Text or frames to video with audio
ComfyUI's template for Text or frames to video with audio: INT8 ConvRot, pruned diffusion model, NVFP4 AWQ text encoder. Stages: text encoder 15.7 GB, diffusion 26.4 GB. Download 42.0 GB.
Computed 3 November 2026References to video with audio
ComfyUI's template for References to video with audio: INT8 ConvRot, pruned diffusion model, NVFP4 AWQ text encoder. Stages: text encoder 15.7 GB, diffusion 26.4 GB. Download 42.0 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
Text or frames to video with audio
On-demand prices before tax, read 19 October 2026.
References to video with audio
On-demand prices before tax, read 19 October 2026.
About the model
Pika H3 Flex is Pika Labs' open video model with sound. Pika Labs' post-trained MiniMax H3 from October 2026: two LoRA adapters, one for text and first-and-last-frame video and one for up to 12 references, that run on the MiniMax H3 base and generate video with synchronised audio in six sampling steps. There is no new architecture, so it needs what MiniMax H3 needs, plus the adapter.
Text or frames to video with audio: 33B · References to video with audio: 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
- Six sampling steps, against 20 to 50 for the undistilled H3, with Pika's schedule built into ComfyUI's three templates.
- Up to 12 references in one generation: 9 images, 3 videos and 3 audio clips.
- Stereo sound, from dialogue to music, made in the same pass as the picture.
- With MiniMax H3 already set up, the adapter is the only new download.
Watch out for
- Underneath it is MiniMax H3, under MiniMax's licence, which does not apply in the United States, the European Union, the United Kingdom or South Korea. Pika's card does not say so; the licence text does.
- Keep the sigma list and the adapter strength as the templates set them, and add no turbo or Lightning LoRA: the adapter is already distilled, and a second one breaks the output.
- Pika's own memory figures are for its CLI package on an H100 and cover sampling only; the text encoder and VAEs come on top.
- The text-to-video template needs ComfyUI 0.43.2 or later.
Every Pika H3 Flex 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 | 66.28 GB | Text or frames to video with audio | minimax_h3_fl2va_bf16.safetensors |
| Diffusion model | INT8 ConvRot | 34.04 GB | Text or frames to video with audio | minimax_h3_fl2va_int8_convrot.safetensors |
| Diffusion model | BF16, pruned | 40.23 GB | Text or frames to video with audio | minimax_h3_fl2va_pruned_bf16.safetensors |
| Diffusion model | FP8 scaled, pruned | 20.96 GB | Text or frames to video with audio | minimax_h3_fl2va_pruned_fp8_scaled.safetensors |
| Diffusion model | INT8 ConvRot, pruned · template | 20.97 GB | Text or frames to video with audio | minimax_h3_fl2va_pruned_int8_convrot.safetensors |
| Diffusion model | W6A8, pruned | 15.98 GB | Text or frames to video with audio | minimax_h3_fl2va_pruned_w6a8.safetensors |
| Diffusion model | BF16 | 66.28 GB | References to video with audio | minimax_h3_ref2va_bf16.safetensors |
| Diffusion model | INT8 ConvRot | 34.04 GB | References to video with audio | minimax_h3_ref2va_int8_convrot.safetensors |
| Diffusion model | BF16, pruned | 40.23 GB | References to video with audio | minimax_h3_ref2va_pruned_bf16.safetensors |
| Diffusion model | FP8 scaled, pruned | 20.96 GB | References to video with audio | minimax_h3_ref2va_pruned_fp8_scaled.safetensors |
| Diffusion model | INT8 ConvRot, pruned · template | 20.97 GB | References to video with audio | minimax_h3_ref2va_pruned_int8_convrot.safetensors |
| Diffusion model | W6A8, pruned | 15.98 GB | References to video with audio | minimax_h3_ref2va_pruned_w6a8.safetensors |
| LoRA | BF16 · template | 1.96 GB | Text or frames to video with audio | pika_h3_flex_fl2va_v1_bf16.safetensors |
| LoRA | BF16 · template | 1.96 GB | References to video with audio | pika_h3_flex_ref2va_v1_bf16.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 Pika Labs says it needs
- "Default 480p/5s sampling allocated approximately 66–69 GiB in representative tests." (Pika's own CLI package on an H100, not ComfyUI; sampling only, with the text encoder and VAEs on top, source)
- "Representative 768p/15s sampling allocated approximately 29–32 GiB on GPU, with observed host memory near 94 GiB including page cache." (Pika's own CLI package with CPU offload, not ComfyUI, 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 Pika H3 Flex
No trainer we checked publishes a VRAM figure for training Pika H3 Flex yet (kohya's sd-scripts and musubi-tuner, ostris' ai-toolkit, diffusion-pipe, and Pika Labs' own repo). When one does, it goes here.
Questions
- How much VRAM does Pika H3 Flex need?
- In ComfyUI's own workflow for Text or frames to video with audio, the largest stage is the diffusion stage at 26.4 GB, and the whole workflow downloads 42.0 GB. Working memory for generation comes on top and grows with resolution and length. Pika Labs itself says: "Default 480p/5s sampling allocated approximately 66–69 GiB in representative tests." (Pika's own CLI package on an H100, not ComfyUI; sampling only, with the text encoder and VAEs on top).
- Can I run Pika H3 Flex on 8 GB of VRAM?
- Not fully. The smallest file here, W6A8, pruned at 16.0 GB, needs 21.4 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 Pika H3 Flex on 12 GB of VRAM?
- Not fully. The smallest file here, W6A8, pruned at 16.0 GB, needs 21.4 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 Pika H3 Flex on 16 GB of VRAM?
- Not fully. The smallest file here, W6A8, pruned at 16.0 GB, needs 21.4 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 Pika H3 Flex file should I download?
- The highest precision that loads fully on your card. On a 24 GB card that is the W6A8, pruned file (16.0 GB). 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.
- Is Pika H3 Flex a new model or a LoRA?
- Two LoRAs. Pika post-trained MiniMax H3 and ships the result as adapters that ComfyUI loads with its LoRA loader on top of the MiniMax H3 base: FL2VA for text and frames, Ref2VA for references. You need the H3 files and the adapter for the job, nothing else, and the licence is MiniMax H3's, territory exclusions included.
- Files and sizes: Comfy-Org/MiniMax-H3 on Hugging Face, read from the Hugging Face API. Template files: ComfyUI's Pika H3 Flex guide.
- Model and licence: Pika-Labs/Pika-H3-Flex 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.