Antirez近期发布了MiniMax H3推理引擎在Mac电脑上的原生实现项目[1]。该项目为Apple Silicon芯片提供了MiniMax-H3文生视频与音频模型的优化推理，支持提示词条件、首尾帧条件以及参考图像、视频、音频等Ref2VA功能[1]。
通过Metal GPU加速与多项优化技术，该实现在M5 Max芯片上实现了显著性能提升[1]。在512x512分辨率、22帧、20步配置下，渲染时间从39.13秒优化至28.06秒，性能提升达28.3%[1]。Int8量化路径进一步优化了完整50层、19次转换的512x512渲染性能，从36.30秒降低到19.32秒[1]。支持--token-reduction模式在512基准测试中可将16.69秒进一步降低到12.60秒，边际提升达24.5%[1]。
该项目还支持交互式会话、Metal编译时优化、内存对齐、图缓存等多项技术优化[1]。测试结果显示四步去噪在fox测试中相对SSIM达0.556（对比29步参考），生成速度约3.5秒[1]。
Antirez has published a native implementation of the MiniMax H3 inference engine optimized for Mac computers equipped with Apple Silicon [1]. The project enables efficient inference of the MiniMax-H3 text-to-video and audio generation model directly on devices such as the M5 Max chip, leveraging Metal GPU acceleration and several optimization techniques [1].
The implementation supports a range of generative capabilities, including text-to-video synthesis, audio generation, first and last frame conditioning, and a Ref2VA feature that accepts reference images, videos, or audio as input [1]. On the M5 Max processor, a baseline configuration of 512x512 resolution with 22 frames and 20 inference steps was accelerated from 39.13 seconds to 28.06 seconds, representing a 28.3 percent performance improvement [1]. When using int8 quantization, the same hardware rendered a full 50-layer, 19-conversion sequence at 512x512 resolution in 19.32 seconds, down from 36.30 seconds [1]. The engine further supports a token-reduction mode that achieves an additional 24.5 percent marginal speedup in 512-benchmark tests, reducing inference time from 16.69 seconds to 12.60 seconds [1].
Performance evaluations indicate that four-step denoising produces outputs with a relative SSIM of 0.556 compared to a 29-step reference on the fox test, with generation taking approximately 3.5 seconds on the M5 Max [1]. The implementation incorporates interactive session support, Metal compilation-time optimizations, memory alignment techniques, and graph caching to enhance overall efficiency [1].