MiniMax-H3-Image lands in vLLM-Omni: 2048² generation in 0.88s on 4×H200
vllm_project · x · 2026-11-10
The vLLM project shipped an official Docker image for MiniMax-H3-Image, supporting NVIDIA (H200, RTX PRO 6000 Blackwell) and AMD (Instinct MI350X) accelerators.
Measured performance (2048×2048)
- 4× B300 (BF16, SP4): 0.63s generation, 0.85s one-reference editing end-to-end
- 4× H200: 0.88s generation, 1.35s editing
- 4× RTX PRO 6000: 1.6s generation, 2.5s editing
Architecture: MiniMax-H3 DiT (33.1B) with a Qwen3-VL text encoder (33.4B), a native 2D-decoder image VAE (0.58B), six CFG-distilled steps per image, up to nine reference images; requests with references are edits automatically. Plan 127GB disk; per-hardware configs in the vLLM Recipes repo.
Related event: vLLM-Omni and MiniMax-H3-Image speed up image generation(2 posts)→
More from Infra
- FlashAttention co-creator Dan Fu on kernels, inference economics, and owning vs leasing GPUs — realDanFu · 2026-11-11
- Self-hosting agent sandboxes on AWS costs a quarter of Daytona's list price — KlausCodes · 2026-11-11
- The Jensen Huang Affair Is the Least Interesting Part: Tracing the $4.8B Nvidia-to-Megaspeed GPU Story — imdigitalashish · 2026-11-11
- SoftBank posts $7B quarterly loss after turning $2B Intel stake into $10.5B and hedging it — realsohamparekh · 2026-11-11
- Email API for agents: appending RFC 5322 copies to Sent via IMAP to fix missing outbound mail — kumard3 · 2026-11-11
- Flide: open-source inference engine with exact state forking claims up to 9.5x speedup over vLLM — _akhaliq · 2026-11-11