推論系統
vLLM-Omni Publishes Four-GPU Rubin Benchmarks for H3-Image: 4K Low Latency and High Throughput Require Different Configurations
On November 12, the vLLM-Omni team published deployment and performance tests for MiniMax H3-Image on Vera Rubin. A four-GPU, short-prompt test reached a median latency of about 0.87 seconds. Peak throughput of 6,115 images per hour came from a different replica configuration, so engineering teams should evaluate interactive speed and service capacity separately.

On November 12, the vLLM-Omni team published inference tests for MiniMax H3-Image on NVIDIA Vera Rubin, bringing native 4096×4096 image generation into sub-second territory. The official short-prompt test used four GPUs to process the same request in parallel, achieving a median latency of about 0.87 seconds. This measurement reflects a specific serving configuration and is not an independently verified, general performance guarantee. Official benchmark
H3-Image uses a single checkpoint for text-to-image generation and reference-image editing. It uses six sampling steps and can generate images of about 16 megapixels without an additional upscaler. The model card lists a 33.1-billion-parameter dense diffusion Transformer, along with a text encoder and an image VAE. Deployment costs therefore depend on more than the diffusion model weights: the encoding and decoding stages must also be included. Model card
The optimizations cover three paths: MXFP8 to accelerate eligible DiT linear layers, Ulysses sequence parallelism to distribute computation across four GPUs, and CUDA Graphs with multiple CUDA streams to reduce VAE decoding overhead. All six denoising steps still run, and the text encoder and VAE retain their respective configured precisions, so the entire pipeline should not be characterized as low-precision inference. Implementation details
The throughput figure is the easiest to misread for capacity planning: 6,115 images per hour came from a configuration with four single-GPU replicas and a concurrency level of eight. Its median latency was 4.411 seconds, making it a different operating point from the sub-second result. For engineers, this means interactive editing and batch image generation should use separately selected parallelization strategies; the two best figures cannot be combined into a single service commitment. Throughput benchmark
Before deployment, check the local request parameters: the model card specifies six sampling steps and a guidance scale of one, and does not support negative prompts. Areas to track include how prompt length, the number of reference images, and weight offloading affect latency. Retesting with your own traffic for tail latency and image quality is necessary to assess the deployment benefits. Parameter limits