This page compares only what MiniMax documents. We do not publish quality benchmarks against other image models; none are cited here.
Standard vs Highspeed (API)
| MiniMax-H3-Image | MiniMax-H3-Image-Highspeed | |
|---|---|---|
| Parameters and limits | Identical per MiniMax docs; switch with the model field | |
| Input text / M tokens | $2.30 | $2.30 |
| Input image / M tokens | $1.50 | $3.00 |
| Output image / M tokens | $3.80 | $7.60 |
| Example 1K text-to-image | $0.02179 | $0.03529 |
| Example 4K text-to-image (list) | $0.06879 | $0.13001 |
| H3-Image-Context-IR endpoint | Supported | Not supported (MiniMax-H3-Image only) |
MiniMax does not publish a latency figure for Highspeed in the docs we checked; the name is the only speed claim.
MiniMax-H3-Image vs MiniMax-H3
| MiniMax-H3-Image | MiniMax-H3 | |
|---|---|---|
| Output | Still images, ~16 MP | Video with audio (H3 family) |
| Architecture | H3 DiT with the audio layers removed | H3-Base design |
| Text encoder | Same MiniMax-H3 (Qwen3-VL) text encoder, unchanged; H3-Image loads it from the MiniMax-H3 repo | |
| Sampling | Distilled to 6 steps, no CFG | See MiniMax-H3 docs |
Practical upshot from the model card: if you already have MiniMax-H3 downloaded, its text_encoder/ is reused and not downloaded again.
MiniMax H3 Image vs Qwen-Image 2.1
Both are open-weights image models that generate and edit in one checkpoint. Facts from each model's own repository; no quality ranking.
| MiniMax-H3-Image | Qwen-Image 2.1 | |
|---|---|---|
| Released | Nov 10, 2026 | Sep 20, 2026 (license date) |
| Generation model size | 33.1B-parameter DiT, 50 layers | 7B visual generation component, 32 single-stream DiT layers |
| Native resolution | ~16 MP (4096×4096) | 2K (2048×2048 default) |
| Default steps | 6 (fixed), no CFG | 40 (Turbo checkpoint: 8) |
| Reference images | Up to 16 | Up to 10, per Oxen.ai's model docs (third party) |
| License | Community license: royalty-free in the Applicable Territory (excludes US, EU, UK, Korea); written authorization above $1M revenue or for hosted services | Qwen Research License: non-commercial research and evaluation only; commercial use needs a separate license |
Open weights vs API
| Self-hosted weights | Official API | |
|---|---|---|
| Cost | Your GPUs; ~135 GB BF16 download | Per token |
| Prompt enhancement | Optional, via MiniMax-H3-Image-PE | On by default |
| Max reference images | 16 | |
| Size limits | ~4–20 MP effective | 655,360–20,000,000 px, aspect 1:3–3:1 |
| Terms | Community license (territory-restricted) | MiniMax platform terms |