Yi Sun

[MiniMax (official) @MiniMax_AI]

More reference control. Fewer denoising steps. More places to run H3.

The latest community contributions span 8-step reference-conditioned video + audio, 4-step distillation, sparse attention, and reference-token compression - alongside serving and workflow integrations across GPU servers, TPUs, and Apple Silicon.

What matters is what this enables: more affordable experimentation, finer creative control, and more choices about where and how video generation runs. Open models become more useful when the community can improve the methods, adapt the runtime, and build workflows around them.

Credit to @haoailab @yisun0618 @lanfeng491270 @guocheng_qian @Alissonerdx @ddalcu @tazmannner @HaochengXiUCB @huggingface @Jhuaplin, and the PTD, DMAD, SparkH3, TorchTPU and FlashML teams - including Zhengming Yu and zechengtang on GitHub.

11 additions to the H3 ecosystem index. Thank you to everyone building, testing, and sharing.

https://t.co/F1S5IjUNi1

[Will Lin @wlsaidhi]

FastVideo's @MiniMax_AI FastH3 Omni model is now available on @huggingface spaces! Try it! Up to 9 reference images, 3 videos, and 3 audio clips!

Thanks to @multimodalart for adding support!

https://t.co/Ul46IHw3Ov

[Hao AI Lab @haoailab]

(1/7) FastVideo FastH3 Omni goes local on DGX Spark, RTX GPUs, and Apple MLX 🚀

- FastH3 Omni, a 5 s 480p clip from a photo and a prompt takes ~3 min. On two Sparks, ~2 min
- FastH3 T2AV is 52.6 s on one Spark, 30.7 s on two Sparks

Blog: https://t.co/FPomdwyJ4t
Cookbook: https://t.co/1mHbHkAyp9
HF 🤗 https://t.co/pMeQUmIm2d

@NVIDIARTXSpark @NVIDIAAI @MiniMax_AI

Every clip in this video was generated on NVIDIA DGX Spark. 8-step NVFP4 checkpoint.

[Jiao Dong @jd_markovchain]

Since day one at Nuva Lab, our core mission has been democratizing creation, using new AI productivity to amplify creativity. For years it was a long, painful wait as we worked our way from text to images to video. We suffered through weak models, skepticism, ai slop hate, closed ecosystems, broken tools, and unit economics too broken to ship at scale.

But, the pieces are finally coming together now. Creative agents are capable enough, open-source video models crossed the quality tipping point, and the open source ecosystem start to wrk. I had pretty much lost interest in grinding on infra, yet I found myself picking my old infra and open-source habits back up to help push the industry forward. Someone had to do it, we wanted to, and we happened to know how to make it happen.

Today, with the FastVideo team, we open-sourced FastH3 Omni as promised after polishing it for a long time. Post-trained from MiniMax H3, it runs 5.6–11.9x faster than the original on 8 NVIDIA GB200 GPUs, and you can deploy it on consumer GPUs. A lot of passionate work went into this behind the scenes across data, training, algorithms, inference, compute, and partnerships, alongside dealing with some bizarre closed-source companies built on open source that tried to trash our work.

I made this video in two afternoons using FastH3 Omni paired with the creative agent we have built video with for a long time. I made it shot by shot in a fully AI-native way, without manual editing or wiring up node-graph spaghetti workflows. For the first time, I ran the exact creation method I imagined years ago without any closed ecosystem limits. So at the end of the video, I put my own Twitter handle in the credits as "Transmedia Director."

This video has plenty of flaws. Still, it hits me hard seeing that one person can finally make something like this, with this method, this model, and at this cost. Maybe our work has changed the world a little bit and brighten up humanity.

Behind this release, there's one group of people building open-source video frameworks year after year, and another group of folks that learned and worked with indie creators of all sizes to top IP studios, and even sometimes get by the days by counting how many fewer people online dmed me "I wish your AI slop die".

Because open source must win.

[Yi Sun @yisun0618]

UniServe supports FastH3 Omni on Day 0!

FastH3 Omni produces high-quality output with fast generation, and it reaches realtime speed in some settings. Our team had a lot of fun building with it internally.

https://t.co/jGaVGzvUav

[Hao AI Lab @haoailab]

As promised, FastVideo open-sources @MiniMax_AI FastH3 Omni, training code included! 🚀

- Real-time gen @ 12x faster on NVIDIA Blackwell
- NVFP4/FP8 for Spark + RTX!
- ComfyUI day-0

Collab with @NVIDIAAI , Nuva Lab @nuvalab and Reactor @reactorworld

Open source must win! ♥️

[Hao AI Lab @haoailab]

(1/6) The all-new FastVideo FastH3-QAD-v2 and the all-new H3 NVFP4 VAE decoders are here! 🚀🚀

This entire 768p video was generated on two DGX Sparks, using QAD-v2 and both new decoders. 8-step video + audio, QAD weights 💾
@NVIDIARTXSpark @NVIDIAAI @MiniMax_AI

[Yi Sun @yisun0618]

@JiaZhihao @SemiAnalysis_ Maybe TCO as y-axis has covered the difference of numbers of chips?

[Hao AI Lab @haoailab]

GPT and Claude now operate real robot arms. What about open-weight models? 🤔
Introducing ORA: Open Robot-Use Agents.
An open harness powered by @deepseek_ai V4.1-Flash, ORA operates robots from visual input—zero-shot, no VLA—and competes with frontier-model agents.

[reactor @reactorworld]

Join us in conversation with Will Lin from FastVideo @haoailab as we get technical about all things realtime video!

Live from 12:30 PT.

https://t.co/1medhnE5Wm

[Hao AI Lab @haoailab]

FastVideo team excited to announce FastH3-Omni! 🚀🚀
Postrained in collaboration with Nuva Lab @nuvalab and Reactor @reactorworld!

The final FastH3-Omni model will be fully open sourced on October 20, along with training code and full support on DGX spark, and RTX GPUs!

Please checkout @reactorworld's API access, available today!

https://t.co/mjjLcRlz51

[reactor @reactorworld]

Introducing FastH3-Omni.

The fastest and highest-quality open-source version of realtime H3 by @MiniMax_AI, post-trained by FastVideo @haoailab with @nuvalab.

Exclusive preview on the Reactor API until 10.20.2026.

• First and last frame conditioning
• Audio input + strong dialogue
• Up to 3 reference images
• Frontier-level performance

[Tianyin Xu @tianyin_xu]

It may not be a common case, but certain absurd behavior at @linuxplumbers Conference #LPC must stop. In the System Boot and Security MC, my student's talk was rudely interrupted; another student presenter received comments like "useless" and "waste of time" during Q&A.

In our talk, the student was interrupted at Slide #4 with a comment that Daniel Borkmann has implemented a patch that largely solved the problem. I just looked at Borkmann's patch [1], while it's related, it does NOT solve the problem. Our work focuses on lifecycle management with dynamic reappraisal [2], while Borkmann's patch is on load-time signing. (In fact, Daniel Borkmann is known for being kind in the community.) Apparently, the person does not even understand the technical problem we are solving – not surprisingly, a conclusion drawn on Slide #4 is unlikely to be sound.

It's great to have active discussion at conferences and it's perfectly fine to have different opinions. However, being respectful is a basic manner and letting people finish their talks is a basic etiquette. This's not even about being kind; it's about being effective in communication – bashing others' work being "useless" won't lead to productive discussion; interrupting won't help understand the work. The goal of critiques is to help colleagues, not to attack.

In the two cases I mentioned, both presenters were students who were too kind and inexperienced to handle such harsh situations. The right way is to take the discussion offline and continue the presentation for the large audience. 

For MC organizers/chairs: Please moderate talks effectively. Interruptions and irrespectful comments should not be encouraged.

For students: If it ever happens to you, don't feel too bad. It's not your fault; it's a shame on whoever not knowing how to communicate.

// It seems that most #LPC discussions are now on LinkedIn, tho the website points to Twitter. Post it here if some colleagues are watching Twitter.

[1] https://t.co/pMai86ucxZ
[2] https://t.co/rrYCdzygDm

[Yi Sun @yisun0618]

@thsottiaux @poteto comments?

[Hao AI Lab @haoailab]

1/8) FastVideo FastH3 now runs on a single consumer machine: @NVIDIA RTX GPUs 💚, DGX Spark and @Apple Silicon 🍎

FastH3 V2 beats base @MiniMax_AI H3 in 8 steps. A 5 s 480p clip with audio takes 15 s on one RTX 5090.

We also release FastH3 Trim: 4.2× smaller than base H3, and it runs in as little as 8 GB of GPU memory.

Every clip here was generated locally. @NVIDIAAI @NVIDIARTXSpark

[Hao AI Lab @haoailab]

Last week we let Jev play Minecraft. 1,795 rounds later, here's how it works, and it's fully open source 🚀
- 🧬 Prompt evolution, no training: a SOTA LLM agent with human supervision evolves the prompt
- ⚡️ 24 ms per decision on NVIDIA Blackwell, 2.5x faster than a leading inference engine
- 🎮 Jev won 70% of Block UHC rounds against real people

https://t.co/fn6lA1kegt

[Yi Sun @yisun0618]

@bcherny @Yuchenj_UW It’s pretty good. I think it matches exactly codex desktop. The ssh is just better - I don’t need to install codex on remote machine then connect from local any more. Would be better if it looks more “professional” like codex.

[reactor @reactorworld]

Serving FastH3 by @MiniMax_AI in real time is a systems problem as much as a model problem.

UniServe from FastVideo @haoailab cuts end-to-end latency to 14.24s on 4×GB200 (vs 18-20s for other serving stacks), and gets it down to 7.51s across 8×GB200. NVFP4 pushes it even further.

Reactor collaborated on this work, read the full breakdown below.

[Yi Sun @yisun0618]

@wlsaidhi @haoailab @MiniMax_AI @reactorworld @nuvalab Huge team efforts!!!