Announced 2026-07-31 with H3-Base weights following in August, under the MiniMax H3 Community License (not Apache; US/EU/UK/KR users must apply separately). A 33B dense transformer with a Qwen3-VL-32B text encoder, native 768P plus stereo audio. Note that only H3-Base is open — instruction understanding (Context-IR) and the 768P→2K upscale stay on the API. The official deployment example uses 4× 80GB cards, a completely different cost curve from the ¥0.50/sec API row above.
Weights are free, hardware is not: you buy the GPU up front (one-time estimate, may rise) and still pay for power and time. Six of the twelve are Apache 2.0 (HappyHorse-1.0, SANA-WM, Bernini, MAGI-2, LingBot-Video, LongLive 2.0); LTX-2 and LTX-2.5 use community licences and SANA-Video uses the NVIDIA Open Model License. Hardware runs from one RTX 5090 to 8× H100. See all 12 →
The official local-deployment example uses 4× 80GB cards.MiniMax · huggingface.co/MiniMaxAI/MiniMax-H3
| # | Platform / model | Layer | VRAM · setup | One-time hardware |
|---|---|---|---|---|
| — | OSSMiniMax H3-Base ↗MiniMax · 2026-08 · community licence · 33B + 32B encoder | Open source | 4× 80 GBOfficial example: 4× 80GB | ≈ ¥1,000,000 · one-time4× H100 80GB · estimate, may rise · cloud rental is more practical |