Intel Arc Pro B70 Adds Day-One Support for MiniMax H3 AI Video Model

Intel is pushing its Arc Pro B70 GPU deeper into the local AI video generation market by supporting the new MiniMax H3 model. This partnership matters because it allows businesses to run high-quality video generation on-premise without relying on external cloud services. This initiative addresses the increasing need for secure, on-premises AI infrastructure in markets with rigorous data governance requirements.

Intel targets enterprise local AI video generation with new model support

The hardware foundation for this capability is the Intel Arc Pro B70, which ships with 32 Xe cores and 32GB of GDDR6 memory. Intel designed a specific 8-GPU deployment strategy to handle the heavy computational load of the MiniMax H3 model. This configuration uses tensor parallelism for text encoding and sequence parallelism for diffusion modules to bypass the memory limits of a single card.

Specifications

• Architecture: 2nd Gen Xe2
• Xe Cores: 32
• VRAM: 32GB GDDR6
• Memory Bandwidth: 608GB/s
• AI Compute Peak: 367 TOPS

Intel claims this multi-card setup delivers 367 TOPS of AI compute peak performance. The system integrates directly with popular open-source frameworks like PyTorch, vLLM, SGLang, and ComfyUI. This compatibility ensures that developers can deploy the solution out of the box without building custom infrastructure from scratch.

The MiniMax H3 model, developed by Shanghai-based Xiyu Technology, generates 15-second videos in 2K resolution with native stereo sound. Intel China provided day-one adaptation for this model, following its earlier support for the DeepSeek series. This rapid integration highlights Intel's strategy to capture local AI deployment opportunities in China.

We looked at the last Intel Arc Pro B70 update, and several of the same balance and stability themes came up. Intel continues to position the Arc Pro B70 as a viable option for enterprise-grade local AI video generation. The focus remains on providing efficient, on-premise solutions for multimodal models.