vLLM brings MiniMax H3-Image to Vera Rubin for subsecond 4K generation

vLLM says its implementation of MiniMax H3-Image reaches a median response time of 0.87 seconds for native 4096×4096 image generation on four NVIDIA Vera Rubin GPUs. A separate configuration reached 6,115 images per hour, according to the company.
- vLLM reports 0.87-second median native 4K generation on four Vera Rubin GPUs.
- Four single-GPU replicas reached 6,115 images per hour at concurrency 8.
- H3-Image supports text-to-image generation, editing and up to 16 reference images.
- The model uses a Qwen3-VL encoder, a 33.1B dense DiT and an image VAE.
- vLLM projects USD 105.16 per GPU per hour at full billable utilization, before operating costs.
Day one support targets high resolution image generation
vLLM has brought MiniMax H3-Image to NVIDIA Vera Rubin through vLLM-Omni. The open-weight model supports both text-to-image generation and image editing, with one checkpoint handling both tasks in six sampling steps.
H3-Image produces native high-resolution images without an upscaler and supports up to 16 reference images. The serving implementation covers single and multi-GPU deployments, allowing users to run the model on a single RTX GPU or distribute generation across multiple data-center GPUs with configurations for different memory capacities.
The company says it has measured the model across 10 NVIDIA hardware types. It also describes support for AMD Instinct, Intel Arc Pro, Huawei Ascend and Moore Threads hardware through separate implementations and recipes.
Separate latency and throughput configurations
The 0.87-second figure comes from a short-prompt serving benchmark using four GPUs, SP4 and MXFP8. At concurrency 1, vLLM reports median latency of 0.866 seconds and p95 latency of 0.944 seconds. The company says this was 2.21 times faster than GB200 and 1.64 times faster than GB300 in the compared configurations.
That headline result is separate from the cross-hardware request benchmark. For a detailed coffee-product prompt, Rubin averaged 1.09 seconds at 4K. In another comparison, four GB300 GPUs averaged 1.48 seconds for 4K text-to-image generation, while a single RTX 5090 with CPU layerwise offload took 52.29 seconds for the same request.
For throughput, four single-GPU replicas reached 6,115 images per hour at concurrency 8, with 4.411-second median latency. The peak-throughput setup is different from the subsecond latency configuration. vLLM reports that SP4 reached 4,407 images per hour at concurrency 2, while replicas added 38.7% throughput.
Optimization work covers compute, decoding and memory
H3-Image combines a Qwen3-VL encoder, a 33.1B dense DiT and an image VAE. A native 4096×4096 output contributes 16,384 target-image tokens across six denoising steps before VAE decoding. Reference images add encoding work and conditioning tokens.
On Blackwell and Rubin, vLLM uses MXFP8 GEMMs for eligible DiT linear layers. Ulysses SP4 distributes sequence work across four GPUs, while DiT TP1 keeps the weights replicated. AdaLN caching reuses conditioning projections when their inputs and weights match, while all six denoising steps still run.
The VAE decodes independent image tiles, which vLLM distributes across GPUs. CUDA Graph replay and multiple CUDA streams are then used to reduce launch overhead and overlap tile processing. In a BF16 control on four GB300 GPUs, the company says VAE decoding fell from 2.635 seconds to 0.213 seconds, reducing end-to-end latency from approximately 4.0 seconds to 1.6 seconds, or 2.5 times faster.
For smaller GPUs, layerwise CPU offload trades weight-transfer time for GPU memory. The documented configurations include RTX 5090, RTX 5080, RTX 6000D and RTX PRO 6000 deployments.
The reported economics depend on utilization and pricing assumptions
Using the measured rate of 6,115 images per hour, vLLM projects 146,755 images per day for a four-GPU Rubin system. The daily figure is an extrapolation of the measured rate rather than the result of a 24-hour load test.
For its revenue estimate, vLLM uses USD 0.06879 per image, based on MiniMax's official Standard 4K API pricing example for text-to-image generation with no references, before launch discounts. At that assumed sale price, the company projects gross revenue of USD 105.16 per GPU per hour at 100% billable utilization, or USD 78.87 per GPU per hour at 75%.
Those figures are before operating costs, and vLLM says actual hosted-API charges vary with text-token usage. For users and operators, the announcement presents H3-Image as a model that can support both interactive 4K latency and higher-volume serving, while the choice between the two depends on GPU count, memory placement, concurrency and utilization.


