MiniMax H3 recipes¶
FastH3 is two distilled MiniMax-H3 checkpoints. V1 is the four-step launch. Some Hub repo names still say Preview. That name is historical. V1 is a full model, not a demo. V2 is the eight-step checkpoint. More forwards is why V2 is the higher-quality FastH3. The V2 schedule contract is in FastH3 distilled checkpoint schedules. CompactH3 is the 42-block 20B NVFP4 H3 checkpoint for one Blackwell GPU (RTX 5090 or RTX PRO 6000). The FastH3 V1 and V2 recipes are unchanged.
The 42-block pruned checkpoint has an MLX INT8/INT6 conversion and eight-forward T2VA command. It reads fastvideo_inference.json for the trained schedule. The command uses native 832x480 resolution and all requested frames.
Primary focus · Inference
MiniMax H3 recipes
Generate video and audio with H3. Run a server on CUDA, one Blackwell GPU, one DGX Spark, or Apple Silicon MLX to iterate on prompts, or call the pipeline directly from Python.
11 maintained recipesCompare H3 modes and options
Supported modes
CUDA covers T2VA, FL2VA, and Ref2VA on the full checkpoint, plus FastH3 V1 and FastH3 V2, plus CompactH3 NVFP4 on one Blackwell GPU. FastH3 V1 also has a DGX Spark runtime with a 1-Spark or 2-Spark device row. MLX is T2VA only: V1 and V2. Temporal --fast, spatial --fast-spatial, and opt-in VSA are flags on the same MLX script, not extra recipes.
| Mode | CUDA | MLX FastH3 |
|---|---|---|
| T2VA | Full H3, FastH3 V1, FastH3 LoRA, FastH3 V2, CompactH3 NVFP4 | FastH3 V1 or FastH3 V2 after a local DiT conversion |
| FL2VA | Full H3 | Not wired |
| Ref2VA | Full H3 | Not wired |
Temporal --fast | No cookbook recipe | Shorter video denoise, MLX RIFE back to --num-frames, full-duration audio |
| VSA | Trained sparse attention on FastH3 CUDA | Opt-in. Convert with --include-vsa into a new directory such as ./FastH3-MLX-vsa, then pass --vsa. Do not overwrite an existing dense export. |
Spatial --fast-spatial | No cookbook recipe | Denoise and decode at height/width divided by --fast-spatial-scale, then resample. Composes with --fast. |
| Two-pass refine | No cookbook recipe | Not wired |
| DGX Spark | FastH3 V1 on one GB10, or two Sparks with Ray sequence parallel (sp_size=2) over QSFP RoCE. Select NVIDIA DGX Spark, then 1 Spark or 2 Sparks. | Not wired |
| CompactH3 NVFP4 | 42-block 20B checkpoint on one RTX 5090 (32 GB, sequential encoder offload) or one RTX PRO 6000 Blackwell (96 GB, encoder+DiT+VAE resident). SageAttention3 FP4, packed NVFP4 DiT, Comfy int8-convrot VAE. | Not wired |
Pick an H3 recipe and runtime
Choose the result you want, then use a maintained CUDA, Blackwell, DGX Spark, or MLX path. Device claims stay tied to checked-in sources and recorded runs.
Loading recipe details...
Exact device and memory details appear only when a recorded run supports them.
Loading...
- Model
- Loading...
- Workload
- Loading...
- Hardware
- Loading...
- Expected output
- Loading...
Loading...Running this script again starts a new process and reloads the model. To iterate in Python, create the generator once and reuse it for multiple prompts.
Start once, then change prompts in the playground or your app. You can run the server and clients on the same machine.
Prepare the machine
Run from your FastVideo clone in an activated Python environment. See installation requirements.
Download and convert MLX weights once
Skip this if the weights are already prepared. Edit the paths in examples/serving/mlx_fasth3.yaml to use your existing files. Install ffmpeg for video and audio output.
Start the server
Keep this terminal running while you use the playground or API clients.
Check that the server is ready
In another terminal, this returns {"status":"ok"} after startup.
Generate and download a video
Edit a prompt, generate, and watch the result. The playground uses the same server as cURL and your app.
Open after the server is ready. On a remote GPU machine, forward port 8000 to your computer first. This opens a local page, not a hosted demo.
Use cURL or an SDK
Each example submits a job, checks its status, and saves the MP4. The Python and JavaScript examples use OpenAI-compatible clients; no OpenAI account is needed.
This is a local development server without built-in API-key authentication. The client key local is a placeholder. Keep the server on loopback; use an authenticated TLS proxy before exposing it publicly. Run the JavaScript client in your webapp's backend, not in a browser with a private key.
Setup
The generated commands expect a local clone:
git clone https://github.com/hao-ai-lab/FastVideo.git
cd FastVideoUse Configuration for supported Python and CLI settings, Optimizations for attention and memory tradeoffs, and the support matrix for the supported model and optimization surface.
CUDA
UV_TORCH_BACKEND=cu130 uv pip install -e ".[fasth3]"Apple Silicon
uv pip install -e ".[mlx]"Follow the MLX install guide for the extra, ffmpeg, and a clone. Then pick FastH3 V1 or V2 in the builder above.
NVIDIA DGX Spark
UV_TORCH_BACKEND=cu130 uv pip install -e .Follow the DGX Spark install guide for ARM64 CUDA 13. One Spark is a local process. Two Sparks need Ray on the QSFP link:
uv pip install rayBring up the cluster from pairing two Sparks before selecting 2 Sparks in the builder.
Troubleshooting
- The full CUDA H3 examples request four GPUs by default. Their sources do not claim a GPU model or memory minimum.
- The FastH3 CUDA performance profile was measured on four GB200 GPUs. Use its strict profile when exact operation order matters more than the measured performance configuration.
- The MLX source runtime supports T2VA, optional temporal
--fast, optional spatial--fast-spatial, and opt-in VSA on--include-vsacheckpoints. FastH3 V2 MLX converts with--include-vsaand runs eight forwards. FL2VA, Ref2VA, and two-pass refinement are not wired. - GPU count and VAE decode backend are configurable in the builder above for FastH3 CUDA recipes. Only the value shown by default has a recorded run; other supported values are unmeasured here.
- DGX Spark is a runtime on FastH3 V1, not a separate family card. Select NVIDIA DGX Spark, then 1 Spark or 2 Sparks. The CUDA GPU-count knob does not apply to Spark.
- CompactH3 NVFP4 is one Blackwell GPU. RTX 5090 (32 GB) parks the encoder in pinned host RAM. RTX PRO 6000 Blackwell (96 GB) keeps encoder, DiT, and VAE resident. Keep
FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER,FASTVIDEO_FA4=0,FASTVIDEO_VSA_SM100A=0, andFLASHINFER_CUDA_ARCH_LIST=12.0a. On PRO 6000 enable VAE compile and leave DiTinference_torch_compileoff. CompactH3 is a dense prune; do not enableVIDEO_SPARSE_ATTN_H3until a VSA-trained student exists. - GB10 has no FA4 / sm_100a VSA kernel. Keep
FASTVIDEO_FA4=0andFASTVIDEO_VSA_SM100A=0. Legalnum_framesvalues are17n+5, capped at 362 (15.08 s). A 345-frame request on one Spark can OOM. Native 16:9 sizes include 832×480 and 1344×768. - Gated or missing checkpoints: run
huggingface-cli loginand confirm you accepted the model's license on Hugging Face.
Evidence status
Every command, model ID, and flag on this page maps to a checked-in FastVideo source. Recipes marked Verified also have a recorded hardware path in linked FastVideo evidence. The full H3 CUDA examples remain Source-backed where the source records a GPU count but no GPU model or memory requirement. Unlisted hardware is unknown, not unsupported.