The open weights of MiniMax H3 have been surrounded by working infrastructure in two weeks: ComfyUI has seen nodes for chunked generation of long videos, stitching clips with per-clip LoRA, and upscaling to 2560×1440, while video acceleration has gone down two independent paths — a distilled FastH3 with 4 steps instead of 49 and 90% sparse VSA attention, and a two-stage Sol-Engine from NVIDIA with a cheap draft and a light refine. The authors claim 14–28× acceleration on data center GPUs, but on consumer cards, generation remains an offline process measured in tens of minutes per clip. We break down exactly what was released, what it means for the cost per second of video, and where the boundaries of known facts lie.

What happened

On August 27, the FastVideo team (Hao AI Lab @ UCSD, Nuva Lab, NVIDIA FastGen) released FastH3 Preview v1: open weights of a distilled MiniMax H3 on HuggingFace with 4 sampling steps instead of 49 and 90% sparse VSA attention; according to the authors' measurements, a 15-second 768p video is generated in less than 13 seconds on 8×B200, and acceleration on a single B200 reaches 14.38×. The NVIDIA SANA team demonstrated Sol-Engine: a 4-step 896×512 draft on H3+LoRA, followed by a 3-step LTX-refine with Sol-Attn; on a single GB200, a 5-second video with a resolution of 1344×768 is generated in 6.852 seconds (22.2× compared to SGLang), and a 10-second video in 14.931 seconds (27.7×). In parallel, a layer of ComfyUI nodes grew around the model: hradec/ComfyUI-HR-Endless-Sampler (73 stars, created on August 17) samples videos of arbitrary length in chunks with latent continuity and live preview, planning and rewriting the prompts of each chunk via Gemma4 12B QAT; tritant/ComfyUI_MiniMax_H3_Extender in version 2.0 (177 stars, updated on September 1) stitches clips with FL2VA support and per-clip LoRA; dntpi/ComfyUI-Hand-Tie-Clips offers a new clip chain; lisitskyaa/ComfyUI_UltimateSDUpscaleGuider_H3 generates 1504×832 with upscaling to 2560×1440.

Context

Before the weights were opened, MiniMax H3 — a generative video plus audio model — was effectively available via a developer API at a price of $0.08 per second of 768p. The open weights changed the access mode: the model can now be post-trained on your own infrastructure, and two proven acceleration patterns were transferred to video. The first is step distillation: FastH3 reduces the basic 49-step sampling cycle to 4 steps and uses 90% sparse VSA attention. The second is a two-stage scheme from NVIDIA: a cheap draft at low resolution is refined by a light LTX-refine with Sol-Attn. The second part of the context is hardware: the headline figures were obtained on data center B200 and GB200, while local ComfyUI workflows live on consumer cards with 16 GB of memory, where it's a matter of offline production, not interactive generation.

Why this matters for the industry

For builders, the main shift is that MiniMax H3 has stopped being just an API: the open weights have made it a platform for post-training, where value is created by a layer of orchestration and acceleration — distillation, sparse attention, draft+refine schemes, and per-clip adapters. Two independent stacks, FastH3 and Sol-Engine, claim acceleration of more than an order of magnitude on the same hardware; compared to the basic 49-step scenario, this significantly reduces the cost per second of generation on your own hardware, and the price umbrella of closed APIs is under pressure. If the pace is maintained, a second wave of distillates and independent reproductions is likely, which will either confirm or adjust the current figures, and LoRA adapters, sparse attention backends, and two-stage refines may become standard components of open video models — following the scenario by which open LLM stacks matured. We also expect a market of packaged solutions on top of such nodes: normal UX, compatibility managers, and workflow presets.

Why this matters for users

If you make videos in ComfyUI, the MVP pipeline "script — long video — 1440p" can be assembled from ready-made nodes today: hradec's Endless-Sampler for chunked generation of clips of arbitrary length with latent continuity (it requires Gemma4 12B QAT for rewriting chunk prompts), tritant's Extender 2.0 for stitching clips with per-clip LoRA and FL2VA support, and lisitskyaa's UltimateSDUpscaleGuider for upscaling 1504×832 to 2560×1440. FastH3 weights with ready-made LoRA adapters can already be downloaded from HuggingFace, but for them to work, the VSA backend from FastVideo is required — dense attention will not do. Expectations on a 16-gigabyte card should be lowered: according to the node author's measurement on an RTX 4080 Super, a 10-second clip takes 18 minutes of generation plus 38 minutes of upscaling, so for now it's batch production and prototypes, not interactive. Keep compatibility in mind as well: some old workflows for Extender 2.0 require node replacement, and the chunked sampler may conflict with Kijai's fast VAE.

What is still unknown / limitations

The key figures — 14–28× acceleration, less than 13 seconds for a 15-second 768p video on 8×B200, 6.852 seconds and 14.931 seconds on GB200 — are measurements by the authors themselves and do not yet have independent replication. There are also no published quality metrics for the distilled FastH3, and step distillation almost always comes with a quality trade-off, so it is too early to talk about an equivalent result "on the same hardware" without its evaluation. The headline indicators were obtained on data center GPUs, so their transfer to consumer cards and the conclusion about an avalanche collapse of the API-second cost remain an interpretation, not an established fact. Finally, the measurement on the RTX 4080 Super was made by the node author on a single configuration, and the compatibility of the nodes with each other — the VSA backend, node replacement in old workflows, possible conflict with Kijai's fast VAE — is checked separately on each setup.

Sources

Author

Look at AI, editorial team