MiniMax releases open-weight image model: 0.8s per 2.7K image on 4×H200
Xianbao_QIAN · x · 2026-11-10
MiniMax released MiniMax-H3-Image, an open-weight text-to-image and reference-image editing model built on a 33.1B DiT with a Qwen3-VL text encoder. It runs 6 CFG-distilled steps per image, hits 0.8s for a 2720×1536 image on 4×H200 via vLLM-Omni, supports up to 16 reference images, up to 4096×4096 output, ships an open-weight prompt enhancer, and natively supports vLLM and verl RL training on your own compute.
Related event: MiniMax open-sources H3-Image: 16MP 4K images in about 10 seconds(20 posts)→
More from Infra
- The AI Race Is Now an Electricity and Permitting Race, Not Just a Model Race — TansuYegen · 2026-11-11
- Route Anything: model routing delivers 8.82X more quality per cost with <0.5-point loss — rohanpaul_ai · 2026-11-11
- Route Anything: model routing yields 8.82x quality per dollar, 25.5x on coding tasks — rohanpaul_ai · 2026-11-11
- Tesla Megapack 3 ramps at Texas Megafactory with 20+ units on site — XFreeze · 2026-11-11
- Cloudflare ships experimental Agents SDK example running the Codex agent loop on Workers — dinasaur_404 · 2026-11-11
- Bare metal GPU sellers target sub-24-month payback, exposing the compute margin stack — AccBalanced · 2026-11-11