FastH3 V2 and FastH3 Trim are FastVideo's distilled MiniMax H3 models for local video generation with synchronized audio. V2 uses eight sampling steps; Trim is an experimental smaller model. The October 6, 2026 release adds consumer hardware packages, giving users a concrete choice of checkpoints rather than a generic promise of faster video.
This article explains the released documentation and model cards as checked on October 7. H3Vid has not reproduced the GPU benchmarks or integrated these checkpoints into its online generator. The practical question is which package fits your hardware and whether its supported inputs fit your project.
V2 versus Trim: what changes?
The Trim source model card describes a 42-block network derived from H3's 50 transformer blocks. The team also compresses timestep conditioning with a shared rank-16 basis. Trim retains eight-step text-to-video with audio, but removing blocks can reduce quality. The authors recommend V2 when quality matters most.
That makes Trim an option to evaluate for memory pressure or quick drafts. It is not a universal upgrade over every four-step model. A shorter forward pass can still produce a longer total job when the sampler performs more passes. Judge the final file and elapsed time together, especially if you already have a working accelerated workflow.
Keep these names separate from hosted H3 Max and H3 Max Turbo. Our Fast H3 page currently offers H3 Max; the FastVideo checkpoints discussed here are a separate local installation choice.
Choose the package before downloading
| Your route | Package to investigate | Important distinction |
|---|---|---|
| Blackwell GPU with FastVideo | V2 NVFP4 Consumer or Trim NVFP4 | Consumer V2 differs from the data-center package |
| RTX 4090 with FastVideo | V2 FP8 or Trim FP8 | Memory offload can affect total time |
| ComfyUI on RTX 50 series | V2 NVFP4 Comfy or Trim NVFP4 Comfy | Use repackaged single-file checkpoints |
| ComfyUI on an older NVIDIA GPU | Trim INT8 convrot | Check the template and installed runtime |
| Apple Silicon | MLX route | Follow MLX instructions rather than CUDA commands |
The V2 Consumer card packages the transformer, a trimmed NVFP4 text encoder, lightweight video decoder, and audio VAE. It includes fastvideo_inference.json for the trained schedule. Machines with less than 64 GB of system RAM can use its documented AdaLN lookup table to avoid loading 26 GB of timestep projections. That table works only with the shipped schedule; incompatible timesteps raise an error.
The Trim FP8 card describes layerwise offload for smaller memory budgets. Checkpoint size is a disk/weight measurement, not the minimum VRAM needed for the complete process. Text encoding, decoding, host RAM, and activation memory also matter.
ComfyUI: use the eight-step template
The V2 Comfy model card specifies the FastVideo FastH3: Text to Video template in ComfyUI 0.36.0 or later. Its NVFP4 file targets NVIDIA Blackwell hardware. The Trim Comfy card provides NVFP4, FP8, INT8 convrot, and BF16 diffusion files.
Documentation-based setup:
- Update ComfyUI and open that FastVideo template.
- Choose one compatible diffusion file and place it in
models/diffusion_models/. - Install a supported Qwen3-VL text encoder in
models/text_encoders/, plus the required video and audio VAEs inmodels/vae/. - Select the downloaded diffusion file in
UNETLoader. - Keep the prescribed eight steps,
res_multistepsampler,simplescheduler, andMiniMaxH3SigmaShiftvalues of 10 for video and 3 for audio.
These instructions come from the cards, not an installation test by H3Vid. Start with the maintained graph before adding custom acceleration nodes. Record the selected files and actual execution settings, so a disabled LoRA or changed sampler does not make later comparisons misleading. Our base H3 ComfyUI guide explains the wider native workflow.
Read the speed numbers with their boundaries
The release article reports warm-server, prompt-to-MP4 timings with audio at 832 × 480 for five-second clips. These include encoding, denoising, decoding, and export; cold model loading is outside that boundary.
| Author-tested machine | V2 | Trim |
|---|---|---|
| RTX 5090, 32 GB | 14.8 seconds | 13.4 seconds |
| RTX 4090, 24 GB | 54.6 seconds | 43.9 seconds |
| DGX Spark, 128 GB unified | 141.4 seconds | 125.8 seconds |
| M4 Max Mac, 36 GB unified | Not reported | 925.2 seconds |
The 8 GB result uses a memory-capped RTX 4090, not a benchmark of an 8 GB graphics card. The authors explicitly expect actual smaller cards to be slower. Neither the headline nor this table establishes the time for a longer, higher-resolution job on your computer.
Mac support and reference inputs
The FastVideo H3 cookbook separates CUDA and MLX routes. Its documented MLX FastH3 mode is text-to-video-and-audio; first/last-frame and reference modes are not wired there. Full CUDA H3 has broader inputs, but that does not establish the same support in every distilled checkpoint.
Before choosing a model for character consistency, check the exact recipe's input contract. A text-only package may be unsuitable when you need an opening image, a closing frame, or a character reference. A community video demonstrating another graph does not extend the maintained recipe's support matrix.
TensorFold Studio is a separate local runtime
As rechecked on October 22, the former keys Mac TensorFold Studio repository now presents Imagination Studio by keys+Wasserman, powered by TensorFold, formerly Slate AI Studio, with Mac, DGX Spark and an RTX Pro 6000 Blackwell route. It is an app built on TensorFold, not a TensorFold product or a new FastH3 model release.
Its README's timing table still identifies the measured Studio configuration as 2.3. On M5 Ultra 256GB it reports 66 seconds for a five-second 864×480 text-to-video run; the DGX Spark row reports 73 seconds with models already loaded. The old Mac column used a different model and pass count, so it is not a controlled before/after comparison. The M1–M4 rows remain explicitly untested estimates from memory measurements on M5. We have not run this app.
Keep these author measurements separate from FastVideo's release table above and from the Omarchy Linux route. Before installing, select the instructions for your operating system and confirm the actual video engine, input mode, hardware and measured version. A newer app heading does not automatically revalidate its earlier timing table.
A useful first evaluation
Choose one short scene with clear motion and one audible event. Run the maintained configuration, retain the output and settings, and inspect identity, movement, fine detail, and sound alignment. Compare V2 and Trim with the same creative brief; record loading time separately from repeated generation. This is a suggested evaluation method, not a test we have completed.
If Trim saves enough memory to run the scene, that is a useful result even when it is slower than your existing four-step setup. If fine detail suffers, keep V2 for final work. If the task requires reference inputs absent from the selected recipe, choose a compatible full H3 workflow before spending time optimizing its speed.