vLLM-Omni brings Day-0 support for MiniMax H3-Image: native 4K in 0.87s
vllm_project · x · 2026-11-12
The vLLM-Omni team announced Day-0 support for MiniMax's H3-Image, the top open-weight image-editing model on Artificial Analysis, optimized for NVIDIA Vera Rubin.
Key details:
- Native 4096×4096 generation hits 0.87s median latency on four GPUs; a throughput config reaches 6,115 images/hour (146,755/day extrapolated).
- At MiniMax's official 4K API price (USD 0.06879/image), peak gross revenue projects to USD 105.16/GPU/hour at full utilization, USD 78.87 at 75%.
- One checkpoint handles text-to-image and editing in six sampling steps, natively high-res without an upscaler, up to 16 reference images.
- Runs on NVIDIA consumer/data-center GPUs, AMD, Intel, Moore Threads (MUSA), and Huawei Ascend NPUs.
Related event: vLLM-Omni Ships Day-0 Support for MiniMax H3-Image with Strong Benchmarks(5 posts)→
More from Infra
- Musk reportedly betting on Terafab: one plant producing a terawatt of silicon chips per year — XFreeze · 2026-11-12
- Physics-informed learning by IITM professor slashes AI simulation compute costs — ravi_iitm · 2026-11-12
- Precision on Demand: Escalating Weight Bits at Uncertain Decoding Steps — Mohammed Nayeem Khan · 2026-11-12
- MiniMax-H3-Image FP8 fits a single 32GB AMD card: 4K images in 33s — evp-cloud · 2026-11-12
- Energy Tech Draws Nearly $200B in 2025 as Data-Center Loads Test Grid Interconnection Queues — ingliguori · 2026-11-12
- Anthropic boots Claude on AMD MI355X in ~3 days, fueling claims the CUDA moat is eroding — udmrzn · 2026-11-12