MiniMax H3 is an open-weight video generation model from the Chinese AI firm MiniMax, available via API and running locally on consumer hardware. Community and cloud variants have cut generation times sharply, while the H3 Max weights fal said it would publish remain unreleased.
MiniMax H3 arrived as an announced model from Chinese AI firm MiniMax, generating up to 15 seconds of audio-synced video from text, images, or voice at 768p or 2K resolution, with support for multiple simultaneous media inputs. In a blind test by Artificial Analysis, it ranked second for text-to-video, behind Gemini Omni Flash and ahead of Seedance 2.0. It was available via API at announcement, with open weights planned.
By August 9, MiniMax had released H3 as an open video generation model that matched ByteDance's Seedance 2.0 in cloud performance. Its Omni Reference feature accepts up to nine images, three videos, and three audio tracks, and it runs on a local PC with an RTX 4090. ASCII.jp tested it with live-action and anime-style prompts.
On the cloud side, fal developed MiniMax H3 Max, which generates 15-second videos in about 13 seconds while maintaining quality. That service offers 5-second videos free up to five times daily, supports Japanese prompts, and costs $0.08 per second at 768p, with a 50% off sale until September 1.
The fastest local results came from a single RTX 5090 running H3, where a 5-second image-to-video clip dropped from 193.5 seconds to 31.6 seconds over 41 days. That 6.12x gain used a 20-step baseline, the official Turbo 8-step LoRA, Hao AI Lab's FastH3 4-step, a community-converted TaoMate 3-step LoRA, INT8 attention, and sparse attention. H3 Max weights have not been released 28 days after fal said on X that it would publish them.
Synthesized by Yomimono from the cited Yomimono stories below, each itself
sourced, then editorially reviewed. Every
fact links the story it came from.
Facts
- Noted
- Video generation cut to 31.6 seconds on a single RTX 5090 · 2026-10-05
- Noted
- One 5-second i2v clip went from 193.5 seconds to 31.6 seconds in 41 days · 2026-10-05
- Noted
- 6.12x performance gain achieved · 2026-10-05
- Noted
- H3 is a 19.5GB open model · 2026-10-05
- Noted
- an open video generation model that matches ByteDance's Seedance 2.0 in cloud performance · 2026-08-09
- Noted
- Its Omni Reference feature accepts up to nine images, three videos, and three audio tracks · 2026-08-09
- Noted
- it runs on a local PC with an RTX 4090 · 2026-08-09
- Release
- game · 2026-08-03 · global · 2026-08-30
- Release
- game · now · global · 2026-08-09
- Announced
- open weights release (planned soon) · 2026-08-09
Connections
Structured graph also available as JSON at /public/entities/minimax-h3.
CC BY 4.0.
Oct 5
A first-person report on GameBusiness.jp timed an RTX 5090 running MiniMax H3 locally. One 5-second i2v clip went from 193.5 seconds to 31.6 seconds in 41 days, a 6.12x gain, using a 20-step baseline, the official Turbo 8-step LoRA, Hao AI Lab's FastH3 4-step, a community-converted TaoMate 3-step LoRA, INT8 attention, and sparse attention. H3 Max weights have not been released 28 days after fal said on X that it would publish them. Cloud H3 Max Turbo ran at 1.49 seconds on different hardware.
Aug 30
The video generation AI MiniMax H3, released as an open-weight model on August 3, has spawned rapid community updates. The AI cloud service fal developed MiniMax H3 Max, which generates 15-second videos in about 13 seconds while maintaining quality. It offers 5-second videos free up to five times daily, with Japanese prompt support, and costs $0.08 per second at 768p, a 50% off sale until September 1.
Aug 9
Chinese AI firm MiniMax released MiniMax H3, an open video generation model that matches ByteDance's Seedance 2.0 in cloud performance. Its Omni Reference feature accepts up to nine images, three videos, and three audio tracks, and it runs on a local PC with an RTX 4090. ASCII.jp tested it with live-action and anime-style prompts.
Aug 9
Chinese AI firm MiniMax announced MiniMax H3, a video generation model that creates up to 15 seconds of audio-synced video from text, images, or voice. It supports 768p or 2K resolution and accepts multiple media inputs at once. In a blind test by Artificial Analysis, H3 ranked second for text-to-video, behind Gemini Omni Flash and ahead of Seedance 2.0. The model is available via API now, with open weights planned soon.