---
license: apache-2.0
base_model: MiniMaxAI/MiniMax-H3
tags: [lora, minimax-h3, text-to-video, audio-video]
---

# H3 LoRAs

Mirrors of the LoRAs our MiniMax-H3 engine depends on, so production never
depends on an upstream repo staying available.

## `minimax_h3_turbo_v4_step600_ema.safetensors`

Verbatim copy of [`larryvrh/MiniMax-H3-Turbo-Lora`](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora)
(Apache-2.0), file `minimax_h3_turbo_v4_step600_ema.safetensors`. Not modified,
not requantised, not re-keyed.

Few-step audio-video sampling. 518 tensors, bf16, ~744 MB, rank 64 (16 on the
AdaLN projections), applied as `W_eff = W + lora_B @ lora_A` with no alpha term.

Measured on our int8_convrot base (`nyxia/H3`, pinkcherry FL2VA) on an H100,
2026-08-10, 124 frames @ 24 fps:

| | |
|---|---|
| base, 50 steps (49 evals) | 20:40 |
| **+ this LoRA, 7 steps (6 evals)** | **4:24** |

Same seed, same conditioning frame, same prompt; output quality judged equal.
Per-step cost rises ~7% (the low-rank branch runs unmerged, since the base is
int8 and merging would cost a dequantise/requantise round trip).

Keys are in **ComfyUI naming**, so it does NOT load through PEFT
`add_adapter()`; it goes through the translation in
`ai-engines-h3/src/ai_engines_h3/lora.py`.

Upstream also publishes a `v1` line (`..._4step_ema_ckpt850`) which their card
calls friendlier for 4-step heavy motion. Not mirrored yet.