返回
原文
MiniMax (official)@MiniMax_AI
39AI 编辑部评分，满分 100
2026-08-09 08:47· 42分钟前
AI 导读

MiniMax 在 Reddit AMA 中公布 H3 系列开源路线图，并承诺“在 AGI 到来前保持开源”。计划开源 H3-Regenerate-2K 潜在空间 DiT 再生模型、统一文生图与图像编辑模型，并考虑转向 Apache-2.0 许可证。H3 采用 MoBA 风格稀疏注意力，支持 4-NFE/8-NFE 低步长变体，Ref2VA 可续接生成 60 秒长视频。

Thank you to everyone who joined our Reddit AMA! The community energy was incredible, and we loved diving deep into the architecture, workflows, and the future of MiniMax H3.🩵

For those who missed it, here is a full recap of what is shipping next and YES WE WILL KEEP OPEN UNTIL AGI ARRIVES.

🔥 The Open-Source Mission As copyright matters settle, transitioning to an Apache-2.0 license is on the table. We also owe you the technical details - a comprehensive technical report on H3's development is in the works and will be published soon~

🎥 Video & H3-Regenerate-2K We are planning to open-source H3-Regenerate-2K. To be clear, this is a dedicated latent-space DiT regeneration model - not just a base checkpoint rerun and not a pixel upscaler. We are currently tuning its efficiency and quality so you can run it locally.

⚡ Architecture, Speed & Sparse Attention Sparse Attention: Our sparse attention is MoBA-style, train-aware block selection. Expect a relatively conservative reference implementation in the near term with the goal of zero perceptible quality loss. Let's build device-specific speedups together!

Low-Step Variants: The released checkpoint is already CFG-distilled. While we don't have a near-term commitment just yet, a 4-NFE / 8-NFE variant is under active consideration. In the meantime, huge shoutout to the community - the Turbo LoRAs you've built are genuinely fantastic.

🖼️ Unified Image Generation & Editing Single-frame image generation was our one of most-asked topics (193 upvotes!). The answer is YES: we plan to open-source a unified text-to-image and general image-editing model derived directly from the H3 lineage. It is currently in post-training refinement.

⏱️ Pushing Limits: 60-Second Workflows For longer videos, Ref2VA supports continuation (just feed the previous clip as the reference). And to the Redditor who chained together that 60-second workflow - great find! That capability is real and was retained from our pretraining phase~🫡

We can't wait to see what you build next. Let's keep pushing the boundaries of open AI together. 🚀

For details👇

RyanLee📢 Official AMA Announcement The complete MiniMax-H3 development team will hold an Ask-Me-Anything session inside r/StableDiffusion. The main researcher team wi...
产品更新图像生成开源生态视频

来源：MiniMax (official) · x.com

导出 Markdown
返回
39在 X 查看原推
导出 Markdown
MiniMax (official) · @MiniMax_AI · X·2026-08-09 08:47·42分钟前
AI 导读

MiniMax 在 Reddit AMA 中公布 H3 系列开源路线图，并承诺“在 AGI 到来前保持开源”。计划开源 H3-Regenerate-2K 潜在空间 DiT 再生模型、统一文生图与图像编辑模型，并考虑转向 Apache-2.0 许可证。H3 采用 MoBA 风格稀疏注意力，支持 4-NFE/8-NFE 低步长变体，Ref2VA 可续接生成 60 秒长视频。

Thank you to everyone who joined our Reddit AMA! The community energy was incredible, and we loved diving deep into the architecture, workflows, and the future of MiniMax H3.🩵

For those who missed it, here is a full recap of what is shipping next and YES WE WILL KEEP OPEN UNTIL AGI ARRIVES.

🔥 The Open-Source Mission As copyright matters settle, transitioning to an Apache-2.0 license is on the table. We also owe you the technical details - a comprehensive technical report on H3's development is in the works and will be published soon~

🎥 Video & H3-Regenerate-2K We are planning to open-source H3-Regenerate-2K. To be clear, this is a dedicated latent-space DiT regeneration model - not just a base checkpoint rerun and not a pixel upscaler. We are currently tuning its efficiency and quality so you can run it locally.

⚡ Architecture, Speed & Sparse Attention Sparse Attention: Our sparse attention is MoBA-style, train-aware block selection. Expect a relatively conservative reference implementation in the near term with the goal of zero perceptible quality loss. Let's build device-specific speedups together!

Low-Step Variants: The released checkpoint is already CFG-distilled. While we don't have a near-term commitment just yet, a 4-NFE / 8-NFE variant is under active consideration. In the meantime, huge shoutout to the community - the Turbo LoRAs you've built are genuinely fantastic.

🖼️ Unified Image Generation & Editing Single-frame image generation was our one of most-asked topics (193 upvotes!). The answer is YES: we plan to open-source a unified text-to-image and general image-editing model derived directly from the H3 lineage. It is currently in post-training refinement.

⏱️ Pushing Limits: 60-Second Workflows For longer videos, Ref2VA supports continuation (just feed the previous clip as the reference). And to the Redditor who chained together that 60-second workflow - great find! That capability is real and was retained from our pretraining phase~🫡

We can't wait to see what you build next. Let's keep pushing the boundaries of open AI together. 🚀

For details👇

RyanLee📢 Official AMA Announcement The complete MiniMax-H3 development team will hold an Ask-Me-Anything session inside r/StableDiffusion. The main researcher team wi...

来源：MiniMax (official)· x.com

产品更新图像生成开源生态视频