Pika Open-Sources H3 Flex Video Model: 15 Seconds With Audio in 6 Steps

Pika has released the weights of H3 Flex, a fine-tuned version of the MiniMax H3 video model, which generates clips with audio in just 6 denoising steps instead of the usual 50. A diffusion video model builds frames from noise: at each step it "cleans up" the image a little, and each such step means a full pass through a heavy neural network. With 6 steps rather than 50, you need roughly eight times fewer passes. That is where the speed comes from: according to the documentation, 5 seconds of video at 1344×768 resolution takes about a minute on a single GPU. The documentation does not say which GPU was used.

What the model can do

H3 Flex generates clips from 5 to 15 seconds long at 24 frames per second. It supports 11 aspect ratios, from widescreen 21:9 to vertical 9:21. Audio is generated together with the picture: 32 kHz, stereo, all delivered in a single MP4 (H.264 video, AAC audio), so there is no need to stitch tracks together. Prompts can be up to 8000 characters long.

The repository contains two LoRAs, i.e. compact add-on modules on top of the base MiniMax H3 model:

• FL2VA takes text plus a first and/or last frame and produces video with audio from them.
• Ref2VA works with references: up to 9 images, 3 videos and 3 audio recordings, but no more than 12 files in total. Each video and audio file must be 2–15 seconds long, and the total duration for each type must not exceed 15 seconds.

Both LoRAs have rank and alpha of 128, and the original weights are in FP32. For ComfyUI there are separate BF16 versions, 1.9 GB each.

How much it weighs and what you need to run it at home

The full set of weights is heavy, about 210 GB:

• H3 transformer: about 62 GB;
• Qwen3-VL 32B text encoder: another 62 GB or so;
• video VAE (the module that encodes and decodes frames): about 10 GB;
• audio VAE: 0.6 GB.

That is too much for a consumer GPU, so the realistic route is through ComfyUI. It pulls in a pruned int8 transformer (part of the weights removed from the model, the rest compressed to 8 bits), the Qwen3-VL encoder in NVFP4 format, and both VAEs.

Pika has not published official VRAM requirements. According to the author of a breakdown in the Telegram channel Neurogen, the working level is 24 GB of VRAM. On 16 GB the model will run with blocks offloaded to system RAM, but slowly. On 12 GB, things go very badly. 32 GB of RAM or more is advisable, because ComfyUI loads and unloads blocks as it works.

You need a recent version of ComfyUI. The text-to-video template requires 0.43.2+, while for Ref2VA and FL2VA the stated version is 0.30.0+. The path is: Template Library → Video → video_pika_h3_flex_t2v.

Resolution and cloud pricing

Natively, the model outputs only 480p and 768p. The 1080p, 2K and 4K options are available only in the Pika cloud, and they are a single upscaling pass from 768p, not true high-resolution generation. If you need genuine 4K detail, don't expect it from this mode.

In the API, you pay per second of finished video, and only for successful generations:

• 480p: $0.01;
• 768p: $0.02;
• 1080p: $0.035;
• 2K: $0.04;
• 4K: $0.08.

Reference mode costs more: $0.03, $0.06, $0.075, $0.08 and $0.10 for the same resolutions, in order. For comparison, 15 seconds at 768p in standard mode costs $0.30.

You can try it without installing anything: the model card on Hugging Face has a free demo Space. It runs on an A10G, so queues are possible.

What this means in practice

Open weights for a video model with built-in audio and multi-reference input are rare. Six steps instead of fifty make local generation tolerable in terms of time, especially on 24 GB. But the speed figures and memory requirements are still approximate: Pika did not name the GPU used for the measurement, and the VRAM estimates come from the author of the breakdown, not from the company. The source did not measure quality against a full 50 steps, so before putting the model into production, you should run it on your own tasks. The good news for studios and enthusiasts is that you can test it through the API for pennies, or in ComfyUI on your own hardware.