vLLM-Omni Reports Sub-Second 4K Generation for MiniMax H3-Image on NVIDIA Vera Rubin

The vLLM-Omni team published a technical blog post on November 12, 2026, announcing "Day-0" support for MiniMax's newly released open-weight image model, H3-Image, and providing detailed benchmarks on NVIDIA's next-generation Vera Rubin architecture GPUs. According to the vendor's self-reported data, using a short-prompt serving benchmark on a system with four GB300 GPUs, the median latency for generating native 4096×4096 resolution images reached 0.87 seconds.

Beyond low latency, vLLM-Omni demonstrated high-throughput configurations: by processing images in parallel across four single-GPU replicas, the system delivered 6,115 images per hour. This performance was achieved through several optimizations, including MXFP8 quantization to accelerate DiT linear layers, Ulysses sequence parallelism (SP4), and VAE decoding improvements combining CUDA Graphs and stream processing. For instance, on four GB300 GPUs, VAE decoding time dropped from 2.635 seconds to 0.213 seconds, reducing end-to-end latency from approximately 4.0 seconds to 1.6 seconds.

Based on MiniMax's official Standard 4K API pricing example ($0.06879 per image), vLLM-Omni calculated potential economics. At full billable utilization, peak throughput projects to $105.16 in gross revenue per GPU-hour; at 75% utilization, this figure is approximately $78.87. It is important to note that these estimates do not deduct operating costs, and the daily capacity is an extrapolation of the measured rate.

Support extends beyond NVIDIA data-center GPUs to consumer GPUs (such as RTX 5090/5080, enabling single-card deployment via CPU layerwise offload) and other hardware platforms, including AMD Instinct, Intel Arc Pro, Huawei Ascend NPUs, and Moore Threads MTT S5000. However, all performance data are vendor self-tests without independent third-party reproduction, and economic projections rely on assumed sale prices and idealized utilization conditions.