MiniMax H3 Turbo Merge GGUF (8-step LoRA baked-in)
The MiniMax H3 FL2VA (pruned) DiT with the LightX2V Turbo 8-step v1.0 LoRA baked in.
and
The MiniMax H3 REF2VA (pruned) DiT with the LightX2V Turbo 4-step v1.0 LoRA baked in.
Modes: img2video, txt2video, FL2VA / T2VA / I2VA / REF2VA.
Samples (4B vs 8B Text Encoder)
Both samples use the same turbo Q4_K_M turbo GGUF diffusion model, prompt, and settings, only the text encoder changes (4B vs 8B).
4B Text Encoder + ClipProj (8-steps)
- (~2.3 min) on an RTX 5090 (32 GB).
Components used to generate
| Component | File | ~Size | Download |
|---|---|---|---|
| Transformer | minimax_h3_fl2va_turbo_Q4_K_M.gguf | 11.4 GB | Link |
| 4B Text Encoder | qwen3vl_4b_fp8_scaled.safetensors | 5.24 GB | Link |
| 4B Clip Proj | mmh3-4b-ClipProj-v3-mlp.safetensors | 304 MB | Link |
| Video VAE | minimax_h3_video_vae_fp16.safetensors | 5.21 GB | Link |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors | 605 MB | Link |
8B Text Encoder + ClipProj (8-steps)
- (~2.5 min) on an RTX 5090 (32 GB).
Components used to generate
| Component | File | ~Size | Download |
|---|---|---|---|
| Transformer | minimax_h3_fl2va_turbo_Q4_K_M.gguf | 11.4 GB | Link |
| 8B Text Encoder | qwen3vl_8b_fp8_scaled.safetensors | 10.6 GB | Link |
| 8B Clip Proj | mmh3-8b-ClipProj-v3-mlp.safetensors | 386 MB | Link |
| Video VAE | minimax_h3_video_vae_fp16.safetensors | 5.21 GB | Link |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors | 605 MB | Link |
Sample settings
| Parameter | Value |
|---|---|
| Prompt | cinematic tracking shot of a red sports car on a wet night street, neon reflections, engine rumble and rain on asphalt, shallow depth of field |
| Resolution | 1280 × 720 |
| Frames | 90 frames → ~3.75 s |
| Steps | 8 |
| CFG | 1.0 |
| Seed | 42 |
Image-to-Video (8-steps)
Link to workflow Link
Components used to generate
| Component | File | ~Size | Download |
|---|---|---|---|
| Transformer | minimax_h3_fl2va_turbo_Q4_K_M.gguf | 11.4 GB | Link |
| 8B Text Encoder | qwen3vl_8b_fp8_scaled.safetensors | 10.6 GB | Link |
| 8B Clip Proj | mmh3-8b-ClipProj-v3-mlp.safetensors | 386 MB | Link |
| Video VAE | minimax_h3_video_vae_fp16.safetensors | 5.21 GB | Link |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors | 605 MB | Link |
REF2VA works with up to 9 input images
Components used to generate
| Component | File | ~Size | Download |
|---|---|---|---|
| Transformer | minimax_h3_ref2va_turbo_Q4_K_M.gguf | 11.4 GB | Link |
| 8B Text Encoder | qwen3vl_8b_fp8_scaled.safetensors | 10.6 GB | Link |
| 8B Clip Proj | mmh3-8b-ClipProj-v3-mlp.safetensors | 386 MB | Link |
| Video VAE | minimax_h3_video_vae_fp16.safetensors | 5.21 GB | Link |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors | 605 MB | Link |
⚠ Must use the updated GGUF Loader ⚠
In comfy:
- Open the ComfyUI Manager
- Change the channel to "Channel (remote)"
- and search for comfyui-gguf-loader
Use the turbo workflow Link
Recommended settings (turbo 8-step)
| Parameter | Value | Notes |
|---|---|---|
width × height | 1280 × 720 | 720P max resolution |
num_frames | 124 | 481 max frames |
frame_rate | 24.0 | |
steps / NFE | 8 | 8-steps minimum, 16 steps max |
guidance / CFG | 1.0 | always use 1.0 for turbo variants |
negative_prompt | (empty in samples) | optional |
Sources
| Base DiT | MiniMax H3 FL2VA pruned fp8_scaled — family of MiniMaxAI/MiniMax-H3 |
| Turbo LoRA | LightX2V FL2VA Turbo 8-step v1.0 (ComfyUI) — minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors |
| Turbo weights | lightx2v/Minimax-h3-Turbo |
| Turbo docs / code | ModelTC/Minimax-H3-Turbo · LightX2V |
| Upstream 8-step specs | FL2VA / T2VA · train 544p mixed AR · video/audio shifts 12 / 3 · distill 8 NFE · recommended inference 8 (upstream also lists 4 for this LoRA) |
| This bake | Offline fuse @ strength 1.0 (208/208 keys) → stochastic re-FP8 → GGUF ladder |
| Runtime LoRA | Not required — already fused into these files |
Notes
- Known Limitations Here
- ComfyUI + 4B/8B: matching
mmh3-*-ClipProj-celeb-mlpis REQUIRED for small TEs. - Audio is first-class on H3 — sample mp4s include stereo AAC.
- Frame grid:
num_frames % 17 == 5(90 and 124 are valid). - MiniMax H3 — custom / other terms; read upstream before commercial use.
- LightX2V turbo LoRA — see lightx2v/Minimax-h3-Turbo and ModelTC/Minimax-H3-Turbo.
- Qwen3-VL — Apache-2.0 (Alibaba).
- ClipProj — MIT code (nicolab28/ComfyUI-ClipProj); matrices per NicoLab’s card.