MiniMax H3: Video, Audio and Cost per Accepted Clip

MiniMax-H3-Max is a separately priced hosted variant: 480P output costs $0.05 per second and 768P costs $0.08. Its first two images are free, then $0.074 each; input video costs $0.0553 per second at 480P or $0.143 at 768P. Do not transfer H3’s five-image allowance or reference-video rates to Max. In the open H3 repository, FL2VA and Ref2VA are task checkpoints: choose frame guidance or mixed references, rather than treating each task folder as an unrelated model.

fal Research’s H3 Max is post-trained from H3 and served through fal’s own stack; it is distinct from MiniMax’s API row MiniMax-H3-Max. fal’s August 27 announcement reports a five-second clip in under three seconds, about 35× the official endpoint’s throughput and 15× faster on average than its comparable-quality group. These are fal’s comparisons. The post identifies GB200 NVL72 infrastructure but does not provide the full GPU allocation, sampling settings, sample counts or retry ledger needed for a matched cost test. Use the selected fal endpoint’s quote, without borrowing MiniMax’s rates; the launch-week discount is not a permanent price.

Use minimax/h3-max-turbo/text-to-video or the separate image-to-video endpoint for this fal product. Checked October 9, its output rates are $0.015/$0.024/$0.048 per second at 480p/768p/1080p. fal labels them a 40% promotion ending October 15, followed by $0.025/$0.04/$0.08. A ten-second 768p output is $0.24 at the promotional rate or $0.40 at the announced regular rate, before any additional charges. Keep Turbo’s price and speed evidence separate from the original fal Max and MiniMax’s direct API. These variants are offered through hosted endpoints; no downloadable checkpoint is listed.

FastVideo’s VSA-DataFree Preview v1 is a separate data-free DMD2-distilled checkpoint requiring its VSA-H3 attention backend. Its documented default uses four B200 GPUs. It supports text-to-audio-video; FL2VA and Ref2VA were not distilled. The author warns that motion, detail and audio can fall below base H3. Evaluate those losses alongside the shorter schedule, not from a secondhand RTX 5090 speed number. The checkpoint inherits the H3 Community License.

Wuli’s community checkpoints support text-to-video and first-frame-to-video with audio. Four steps use a LoRA plus video/audio output projections on the official transformer; the six- and eight-step variants contain a full Diffusers transformer. All need official H3 base components. The workflow saves video and audio together as an MP4. H3 Community License applies. Choose the matching workflow and measure memory and accepted-clip time on your hardware; fal’s hosted Turbo price belongs to a different product.

FastVideo V2 uses eight transformer forwards, data-free DMD2 and VSA-H3 at 80% sparsity. Its video scheduler shift is 10, compared with base H3’s 12, and it requires the FastVideo VSA-H3 backend. Documented defaults use four B200 GPUs; the GPU count must divide 56 attention heads. It supports text-to-audio-video, with FL2VA and Ref2VA outside the distilled scope. Motion, fine detail and audio can fall below base H3. H3 Community License applies. Keep these settings separate from Preview v1 and Wuli’s Diffusers workflows.

The author reports a Windows RTX 4060 Laptop with 8GB VRAM and 16GB RAM, using pruned W4A8 FL2VA weights, a Qwen3-VL 4B encoder with ClipProj, Spectrum and low-VRAM offloading. The recommended tested setting is 480 × 864, 158 frames and 20 Euler/Simple steps; the supplied graph defaults to 608 × 352 and 124 frames. It produces video without an audio connection. The author reports roughly 10–11+ minutes for a separate 544 × 960, 158-frame, 20-step run on that laptop. Its exact timing is pending; the recommended 480 × 864 setting has no recorded time. Use the component list and performance notes to budget memory and clip time.

Hao AI Lab’s experimental Trim removes eight of 50 transformer blocks and uses eight denoising steps. Its FP8, 832 × 480, five-second clip with audio took a reported median of 82 seconds on a warm RTX 4090 artificially capped at 8GB GPU memory. The warm timing includes encoding, denoising, decoding and MP4 export, with two runs on each of two prompts. The authors expect a real smaller-memory card to run slower than that capped 4090. Keep the hardware and precision with the timing.

This experimental LoRA removes a selected person and reconstructs the background. Its ComfyUI workflow adds SAM 3.1 masking, a separately cleaned first frame and overlapping generation windows. The adapter still needs the base, encoder, VAEs and segmentation models. Start with a short continuous shot, inspect masks and reroll weak windows; changed background details and shadows affect accepted-edit cost. The supplied workflow produces silent output by default; retaining the original soundtrack is a separate connection/postprocessing step. H3’s license applies; the relay code’s GPL license does not relicense the weights.

The original Union card now recommends 2.0 for new work. This H3 control branch accepts eight controls including edges, depth, pose, sketches and layouts, plus video inpainting. Its roughly 13.5GB branch is additional to base H3 and the encoder. Use the matching 2.0 configuration: the old five-block configuration can silently discard part of the checkpoint. Include control preprocessing, extra memory and retries in your budget. The model card retains the H3 Community License.

Comfy-Org explicitly repackages upstream model files for ComfyUI, including BF16, INT8 ConvRot and FP8 choices, encoders, VAEs and adapters. A smaller transformer file does not eliminate encoder or runtime memory. Match the task, precision and compatible runtime, then measure peak memory and seconds for the same clip. This packaging is not itself a newly trained base model. Its card uses the H3 Community License; keep adapter terms with the upstream license.

A provider channel adds its own model ID, supported inputs, resolutions, queue limits and billing rules. Confirm whether it serves H3, H3-Max or a modified checkpoint. Compare the total for the same references and delivered clip, including minimum duration and retry charges. The provider’s service terms govern that API purchase; access to an endpoint does not grant a separate right to redistribute H3 weights.