These models are meant for research purposes and for testers currently testing out DF. If you are neither, they likely won't be much use to you.
Pure INT4 (Ultra-Low VRAM Experimental)
A pure 4-bit uniform quantization designed for minimal memory consumption.
### 🔬 Model Details & Caveats
* VRAM Footprint: *9.80 GiB** (Ultra-compact, maximum memory headroom).
* Quantization Scheme: Symmetric 4-Bit Linear FastInt4Linear) with Group-128 scaling.
* Adapter Support: Requires loading the external LightX2V 4-step Turbo LoRA minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors) if 4-step generation is desired.
⚠️ *Experimental Note**: In extended runs (e.g. 10-second / 124+ frame sequences), pure INT4 quantization on time-modulation layers may exhibit slight identity drift or rhythmic "zombie sway" artifacts. Use DF-Turbo-W4A8 if artifact-free motion is required.
---
## 🌟 Variant 1: Digital Forge Turbo W4A8 (All-In-One 4-Step) — This has failed. The video side worked and gave hope for audio, but the difference for audio is hard to circumvent. Audio doesn't have independent blocks or lanes, it shares with video. So, very close, but no cigar in the end. Back to the drawing bored. It seems others ran into similar or the same issue. We found block 47 was able to improve quality of audio, but never fully matched bf16 in quality for audio.
---
## 📜 License & Attribution
Base model licensed under the *MiniMax H3 Community License Agreement**, Copyright © 2026 MiniMax.
Powered by *MiniMax H3** & Digital Forge.
Description
Comments (6)
Could you share a workflow for this model? I used the "minimex h3 easy" workflow, but the videos didn't turn out well.
Yeah, I don't use ComfyUI anymore. Different backend and frontend completely. System that uses them isn't public. I have no workflow to share. This will be the case going forward, I use my own proprietary tools these days.
Same question about the workflow it would be nice to have an exemple where it work !
Yeah, I don't use ComfyUI anymore. Different backend and frontend completely. System that uses them isn't public. I have no workflow to share. This will be the case going forward, I use my own proprietary tools these days.
What is the max resolution and number of frames you can push int4 to compared to non-int4 version?
Can you run 1080p on a consumer level card with int4?
With a patched sage attention (fixes a similar issues to LTX with the nans but reduces the speed gain slightly), custom vae, some helper scripts, yes. Max resolution hasn't been tested yet, current testing is limited to 720p for now. Focus is on speed without significant quality loss, the int4 is mainly to reduce whats on the vram to avoid swapping and offloading as much as possible. In theory, should be able to do 1080p fine, but hasn't been tested.