MiniMax H3 gets fast locally: Sol-H3 Studio, HyperFlow 8-step, Ref2VA-VSA

MiniMax H3 is the open video+audio model behind a lot of this week's ComfyUI workflows, and the bottleneck has been speed. Today NVlabs added Sol-H3 in ComfyUI to the sol-engine branch of NVlabs/Sana (first commit Oct 16, announced by @xieenze_jr). It's a full-screen 'Sol-H3 Studio' over ComfyUI: prompt, mode, pipeline, Generate, then a gallery. The model runs in its own worker process, so ComfyUI's packages stay untouched, and you can flip back to the node graph any time.

These are the README's own numbers for one 1344×768, 121-frame clip with audio on an RTX 5090, from request to MP4: stock H3 at 50 steps took 989.0s. The HyperFlow 8-step LoRA took 193.6s. Sol-H3 one-stage (8-step LoRA merged into an FP8 DiT, Sol-Attn, compiled, fast VAE) took 67.2s. Sol-H3 two-stage (H3 draft + LTX refine) took 35.5s. A DGX Spark two-stage package did it in 64.3s, and one-stage 480p in 22.0s. The catch is the requirements: Linux, one RTX 5090 (32 GB), at least 90 GB of host RAM, Python 3.12, PyTorch 2.12/CUDA 13, and FlashAttention 4 + SageAttention 2.2 built for SM120. This is an enthusiast-rig install, not a one-click node.

Two pieces underneath are older but worth knowing. HyperFlow (Video Rebirth, repo created Sep 17) is the 8-step H3 LoRA that Sol-H3 builds on, and it has community ComfyUI ports such as Saganaki22/ComfyUI-Hyperflow. ComfyUI-Ref2VA-VSA by Kablex (Apache-2.0, created Sep 5, resurfaced today by @Forhanvv) targets H3 reference-to-video, meaning character-consistent shots. Its README reports a 5.16s 1344×768 clip in about 72s warm on an RTX 4090 at 4 steps with 75% video sparsity and about 13.5 GB peak VRAM, compared with about 135s for VDN-H3. The post puts native dense H3 at about 650s. The README also says reference tokens are kept out of attention pruning to protect identity.

Our take: if you run H3 locally, today's Sol-H3 numbers are the biggest jump we've seen. But most people should start with the HyperFlow ComfyUI nodes or Ref2VA-VSA, which fit 24 GB cards. All timings above are the authors' claims; we haven't benchmarked them.