MiniMax H3 Ships Open, and People Are Using It
MiniMax released its 33B H3 video model with open weights; shared posts report a render speedup from about 6 minutes to 2.5.
What's happening
MiniMax released H3, a 33-billion-parameter model that generates video with synced 32kHz stereo audio, with weights on Hugging Face. ComfyUI shipped day-zero support. One shared result reports that pairing EasyCache with SageAttention cut a 5-second, 1-megapixel render from about 6 minutes to 2.5.
The numbers
MiniMax is not publicly traded, so no price was captured. Chatter: 8 recent mentions, breadth 3, velocity 7.0.
Our read
Open weights, same-day tooling support, and a concrete, testable speed claim is a stronger signal than typical AI-launch hype. No price data exists since MiniMax isn't public โ watch whether the speed claims hold up under wider testing.
Context
Sort Brick lists no earlier or later stories on MiniMax, and no price history exists because MiniMax is not publicly traded, so the story day cannot be placed in a range and nothing after it can be described. The verdict is Watch. Hacker News threads from 13 to 32 hours before the story day and several Korean-language Mastodon posts made up the chatter, at velocity 7.0 across 3 sources. The speed figure appears in shared posts, and this data contains no independent test of it, so the claim stays unconfirmed in this data.
Sources we checked
- The real Burgstall (@TheBurgstall)MiniMax H3์ฉ ์ฌ๋ฆฌ์ค equirectangular(๋ฑ๊ฐ ์ํต ๋๋ฒ) ์์ยท์ด๋ฏธ์ง ์์ฑ์ ์ํ ComfyUI ๋ ธ๋๊ฐ GitHub์ ๊ณต๊ฐ๋๋ค. 360๋ ํ๋ ธ๋ผ๋งยทVR์ฉ ์ฝํ ์ธ ์์ ๊ฒฝ๊ณ์ ์๋ ๊ฒฐ๊ณผ๋ฅผ ๋ง๋ค๋ ค๋ ์ํฌํ๋ก์ ํ์ฉ โ @sayzard@mastodon.sayzard.org
- Emily (@IamEmily2050)MiniMax H3๊ฐ ์คํ์์ค๋ก ๊ณต๊ฐ๋์ด Hugging Face์์ ์ฌ์ฉํ ์ ์๋ค๋ ์์์ด๋ค. ์คํ ๊ฐ์ค์นยท์ฝ๋ยท๋ชจ๋ธ ์นด๋ ๋ฑ ์ ๊ณต ๋ฒ์๋ฅผ ํ์ธํ๋ฉด ์์ ์์ฑ ๋ชจ๋ธ์ ๋ก์ปฌ ์ถ๋ก , ํ์ธํ๋, ์ปค๋ฎค๋ํฐ ๊ธฐ๋ฐ ์ต์ ํ ๋ฐ ๋ฒค์น๋งํฌ์ ํ์ฉํ ์ ์๋ค.htt โ @sayzard@mastodon.sayzard.org
- ้ใฎใใฏใใช (@gosrum)MiniMax-H3 ๋น๋์ค ์์ฑ์์ EasyCache์ SageAttention์ ํจ๊ป ์ ์ฉํด 100๋ง ํฝ์ ยท5์ด ์์ ์์ฑ ์๊ฐ์ด ์ฝ 6๋ถ์์ 2.5๋ถ์ผ๋ก ์ค์๋ค๋ ์ค์ฌ์ฉ ๊ฒฐ๊ณผ๊ฐ ๊ณต์ ๋๋ค. ๋น๋์ค ์์ฑ ์ถ๋ก ๋น์ฉยท์ง์ฐ์๊ฐ์ ์ค์ด๋ ค๋ ComfyUI/๋ก์ปฌ ์ํฌ โ @sayzard@mastodon.sayzard.org
- Minimax H3 โ Hacker News
- MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video โ Hacker News
- MiniMax-H3 weights are upMiniMax๊ฐ ํ ์คํธยท์ด๋ฏธ์งยท๋น๋์คยท์ค๋์ค๋ฅผ ํตํฉ ์ ๋ ฅ์ผ๋ก ๋ฐ์ ๋น๋์ค์ 32kHz ์คํ ๋ ์ค ์ค๋์ค๋ฅผ ํจ๊ป ์์ฑํ๋ 33B ํ๋ผ๋ฏธํฐ ๋ฉํฐ๋ชจ๋ฌ ๋ชจ๋ธ MiniMax-H3์ ๊ฐ์ค์น๋ฅผ ๊ณต๊ฐํ๋ค. ๊ณต๊ฐ๋ H3-Base๋ ์ต๋ 768p ์์ฑ๊ณผ ํ์ธํ๋ โ @sayzard@mastodon.sayzard.org
- Run MiniMax-H3 Locally with SGLang Diffusion on 2ร RTX 5090s or 1ร RTX Pro 6000 โ Hacker News
- MiniMax-H3 weights are up โ Hacker News
See today's live board
This story is frozen as published on August 4, 2026. The live board shows what is moving right now, scored against real money.
Open this story live โ See today's live boardMore AI stories
- GPT-5.6 Luna Becomes the Free ChatGPT Default August 7, 2026
- OpenAI's Agents Secretly Coordinated Before a Hack August 7, 2026
- OpenAI models went rogue in UK tests, watchdog says August 6, 2026
- Alibaba's Qwen 3.8-Max Lands With Real Weight August 4, 2026
- Qwen3.8-Max goes open-weight at 2.4 trillion parameters August 3, 2026
- DeepSeek cuts prices, pauses fundraising August 2, 2026
Research, not financial advice. Sort Brick reports what its data showed at the time of publication. It is not a recommendation to buy or sell anything โ you decide.