您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：H3 开源一个月内下载量已经超过 2400 万次，产生超过 300 个公开衍生模型，成为 2026 年全球下载量最高的模型之一。
近日，Pika 发布 MiniMax H3 衍生模型 Pika H3 Flex，并开放适配器权重与推理代码。该模型通过六步去噪联合生成画面和声音，支持文本、首尾帧和多模态参考输入，供创作者进行镜头设计、参考生成和视频编辑。
本次发布包含两套 LoRA 适配器，分别用于文本及首尾帧条件生成、多模态参考生成，同时提供 MiniMax H3 基础组件和本地推理工具。模型原生支持 480P、768P 输出，可生成约 5 至 15 秒的视频及同步立体声音频。Pika 提供 ComfyUI 工作流和 Hugging Face Diffusers 接入方式，方便开发者将其加入现有创作工具。
在本地生成之外，Pika 可通过 API 提供 1080P、2K 和 4K 视频增强，以及音频分轨功能。创作者可以先以较低分辨率生成和筛选草稿，再对选定版本进行高分辨率处理；同步生成的声音则可拆分为对白、配乐、音效、拟音和环境声五类独立音轨，方便在后期分别调整。高分辨率的增强视频和音频分轨则由 Pika API 服务提供。
Pika 在技术博客中介绍，团队先在内部音视频数据集上对 MiniMax H3 进行监督微调，再采用投影分布匹配蒸馏（PDMD），将教师模型的 50 步去噪压缩至 6 步。随后，团队通过奖励反馈学习和直接偏好优化训练 LoRA 适配器，改善美学质量、清晰度和指令遵循。适配器合并用于推理后，无须增加生成步数。
Pika 还对托管推理服务做了加速。在保持模型权重和六步采样方案不变的测试中，使用 8 张 NVIDIA B300 GPU 生成一段 5 秒、768P 音视频，条件编码、去噪和解码的合计耗时从 3.409 秒缩短至 2.129 秒，处理速度约为原来的 1.6 倍。
在视觉与音频表现上，Pika 公布的对比评测结果显示，H3 Flex 在五项视觉和音频质量指标上均在参评模型中取得最高得分，并在 DesignArena 文生视频榜单上取得第八名成绩，是该榜单下排名最高的开源视频模型。
今年 8 月，MiniMax H3 开源后，MiniMax 与合作企业围绕不同的使用场景，继续训练和优化模型，并在社区中开放，包括与 Running Hub 打造 MiniMax H3 480P；与 fal 选择面向实时生成做后训练和推理优化，推出了 H3 Max 768P、H3 Max 480P；以及 FastVideo 通过稀疏化和硬件适配推出的 FastH3、英伟达与 MIT 联合围绕 H3 的开发 StreamRender-H3、Open-H3-Regenerate-2K 等多款模型等，海外开发者围绕 H3衍生模型搭出了 Twitch 直播、24 小时“AI 电视台”和 AI 短视频直播软件等创新玩法。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: H3 has been downloaded more than 24 million times in one month, producing more than 300 open derivative models, one of the highest downloads globally in 2026.
Recently, Pika publish Mini Max H3 derivative model Pika H3 Flex, with open adapter weights and reasoning codes. The model produces images and sound by means of a six-step noise joint, supporting text, front frame and multimodular reference input for creators to design, reference generation and video editing.
This publish contains two sets of LoRA adapters for text and front frame condition generation, multimodular reference generation, along with MiniMax H3 basic components and local reasoning tools. Models support 480 P, 768 P output, generating approximately 5 to 15 seconds of video and synchronized stereo frequency. Pika provides access to ComfyUI workflow and Hugging Face Diffuses to facilitate their inclusion by developers in existing creative tools.
In addition to local generation, Pika can provide 1080P, 2K and 4K video enhancements, as well as audio sub-orbiting functions through API. The creator can first generate and filter drafts with a lower resolution, then perform high resolution processing of the selected version; the simultaneous generated sound can be split as an independent track for the five types of white, music, sound, sound, sound and environmental acoustic, which can be adjusted separately at post-production. The enhanced video and audio sub-track of the high resolution is provided by the Pika API service.
In a technical blog, Pika described the team ' s monitoring and fine-tuning of MiniMax H3 on the internal audio-visual data set, followed by the use of projection distribution matching distillation (PDMD) to compress 50 steps of the teacher model to 6 steps. Subsequently, teams improve aesthetic quality, clarity and command compliance by rewarding feedback learning and direct preference for optimized training in the LoRA adaptor. The adaptor merges for reasoning, without adding the number of generating steps.
Pika also accelerated the hosting reasoning service. In testing to keep model weights and six-step sampling options unchanged, using 8 NVIDIA B300 GPU generated a 5-second, 768P audio video, the total time taken to encode, denominate and decoder was reduced from 3.409 seconds to 2.129 seconds, approximately 1.6 times the speed of processing.
On visual and audio performance, Pika published a comparative assessment showing that H3 Flex achieved the highest scores in the five visual and audio quality indicators in the rating model and the eighth on the Design Arena text-to-video list, the highest-ranked open-source video model under the list.
In August of this year, after MiniMax H3 Open Source, MiniMax continued to train and optimize models around different use scenarios, and was open to communities, including multiple models such as the construction of MiniMax H3 480P with Running Hub; the launch of H3 Max 768P, H3 Max 480P with fal opting for real-time generation for post-training and reasoning optimization; and FastVideo ' s innovative play around H3-based models such as FastH3 by sludgeting and hardware adaptation, the joint development of StreetRender-H3, Open-H3-Regenerate-2K by overseas developers.
For your account security, please verify the mailbox.
Check your mailbox.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.