We're open-sourcing H3 HyperFlow. A data-free flow self-distillation technology built on MiniMax (official) H3. No external training data. Significantly reduces inference cost while preserving frontier model quality. Full demo:

HyperFlow delivers: 🎥 Better camera control 🔄 Better consistency 💎 Better material and detail ⚖️ More balanced across capabilities Same foundation. Better results.

HyperFlow also supports @nvidia 's Sol-Attn sparse-attention kernel, delivering up to 13% faster H200 generation than dense attention.

The key: it's data-free. The model generates its own candidate outputs, filters for quality, and self-distills from that. No external data. No human labels. The model teaches itself. This is the self-distillation component of RSI (Recursive Self-Improvement), our training paradigm for video models.

Why data-free matters: no external datasets. No RL infrastructure. No human labels. The model teaches itself, and that's all the loop needs. We proved it on our own base model. HyperFlow proves it on @MiniMax_AI H3.

HyperFlow is open source. Full weights and inference code. Based on @MiniMax_AI H3, under the MiniMax H3 Community License. Code: Weights:

Frontier quality. Lower inference cost. Four specific improvements. And it's open.

@MiniMax_AI what advantage yours has compared to Lightricks 8 and 4 steps loras?

At 8 steps we're seeing parity or better across our eval set, with the biggest visible difference on motion-heavy prompts. 4-step works too but you'll notice fine detail and fast motion degrade first. Eval scripts are in the repo if you want to run the same tests on your own prompts.

@MiniMax_AI Open sourcing HyperFlow is a game changer for efficient AI inference.

@MiniMax_AI Opensource 🔥🔥

@MiniMax_AI Another slow motion problem lol

@MiniMax_AI Could this method be used to create a 2 step lora?

@MiniMax_AI The current release isn’t optimized for such aggressive configurations. You may encounter artifacts when running 2-steps inference directly. If you do experiment with it we'd genuinely like to hear what you find.

@MiniMax_AI The data-free self-distillation approach is the part that really stands out. Better quality, lower inference cost, and no external data or human labels in the loop. Making the full stack open source makes this even more interesting for developers.

@MiniMax_AI And this is how the foundation labs have been doing it for over a year. recursive self distillation.

@MiniMax_AI you're right

round 1: teacher T0 generates a corpus of reasoning traces. Not just answers but full chains of thought, with the dead ends included. Student S1 trains on those traces. Standard distillation so far. Round 2: S1 becomes T1. It generates a new corpus. But here’s the beautiful catch, T1 generates traces that are cleaner than the ones it trained on, because during training it learned which parts of T0’s reasoning were strong and which were noise. The distillation acts as a filter. Signal goes up. Round 3 onward: this compounds. Each generation’s traces are cleaner than the last. The model is not learning new facts about the world. It is learning to think more efficiently about the facts it already has. Compression of cognition. Am I on the right track? 😉

@MiniMax_AI 若抓包结论属实，“加密上传”不等于隐私保护：密钥在云端，服务方仍具备解密能力。关闭开关仍上传更是越过授权边界。企业需要的是可验证的禁传、删除审计和本地密钥，而非一纸隐私条款。

@MiniMax_AI No external data is the interesting part

@MiniMax_AI The H3 HyperFlow was so smooth 🫶🏻

@MiniMax_AI amazing letting the model generate and filter its own training signals is a fascinating loop.

@MiniMax_AI Open-sourcing efficient video inference is a major step forward

@MiniMax_AI Huge move open-sourcing H3 HyperFlow cutting inference cost without external data while keeping frontier quality is massive!

@MiniMax_AI Data-free self-distillation is the interesting part here, since most inference-cost wins lean on a teacher dataset or heavy calibration. Curious whether the quality holds on out-of-domain prompts or only on the H3 distribution it was distilled against.

@MiniMax_AI nice, another route to cheaper inference. local agents are getting crowded lol

@MiniMax_AI Data-free distillation with lower inference costs is a compelling direction, especially if quality really holds up.

@MiniMax_AI Such a amazing