Blog · October 9, 2026
Run MiniMax H3 locally: GPU, VRAM and RAM requirements
MiniMax H3 is open weight, so you can run it on your own hardware. The full model is large: about 124 GB of weights in full precision. Quantized builds run on consumer GPUs, but a 5-second clip can take many minutes, and system RAM becomes the limit.
Model size
How big is MiniMax H3?
| Part | Full precision | Smallest ComfyUI build |
|---|---|---|
| Transformer (33B) | 66.3 GB (BF16) | 16.0 GB (pruned W6A8) |
| Text encoder (Qwen3-VL-32B) | 51.5 GB (BF16) | 15.7 GB (NVFP4-AWQ) |
| Video VAE | 5.2 GB (FP16) | 5.2 GB |
| Audio VAE | 0.6 GB | 0.6 GB |
| Total | 123.6 GB | 42.5 GB |
Sources: Comfy-Org/MiniMax-H3 files and ComfyUI launch post. Read October 9, 2026.
VRAM
MiniMax H3 VRAM and RAM requirements
| GPU memory | Setup | 5 s clip |
|---|---|---|
| 4× 80 GB (H100) | Base H3, BF16, SGLang, 50 steps | 200.1 s at 1344×768 |
| 96 GB (RTX PRO 6000) | FastH3 V2, NVFP4 | 36.5 s at 1344×768 |
| 32 GB (RTX 5090) | FastH3 V2, NVFP4 | 38.6 s at 1344×768 |
| 24 GB (RTX 4090) | FastH3 V2, FP8 | 154.6 s at 1344×768 |
| 16 GB (capped RTX 4090) | FastH3 V2, FP8 | 153.7 s at 1344×768 |
| 12 GB (capped RTX 4090) | FastH3 V2, FP8 | 170.7 s at 768p |
| 8 GB (capped RTX 4090) | FastH3 Trim, FP8 | 82.0 s at 832×480 |
| 128 GB unified (DGX Spark) | FastH3 V2, NVFP4 | 141.4 s at 832×480 |
| 12 GB + 32 GB RAM | Base H3, SGLang consumer recipe, 20 steps | 319–356 s at 864×480 |
Sources: Hyperstack, Hao AI Lab, SGLang cookbook. Read October 9, 2026. The 16 GB, 8 GB and DGX Spark rows are from the same Hao AI Lab post, read October 17, 2026.
FastH3 Trim is an experimental FastVideo build that removes 8 of the 50 transformer blocks. Hao AI Lab reports that both FastH3 V2 and Trim fit in 12 GB, and that Trim fits in 8 GB. The 16, 12 and 8 GB rows cap GPU memory on one RTX 4090. A real card with less memory will be slower (Hao AI Lab, October 6, 2026; read October 17, 2026).
MiniMax gives no official minimum VRAM. Low VRAM moves weights to system RAM. The vLLM-Omni recipe for one RTX 5090 asks for at least 200 GiB of host RAM and advises 384 GiB. Community reports go down to 6–8 GB of VRAM with heavy offload, but a 5-second clip then takes 10 minutes or more.
FastH3-Omni
FastH3-Omni local times: RTX 5090, RTX 3090, DGX Spark and Mac
| Machine (weights) | 1 image | Image + voice | 1 short clip |
|---|---|---|---|
| RTX 5090 (32 GB), NVFP4 QAD | 61.9 s | 63.6 s | 57.2 s |
| 2× DGX Spark, NVFP4 QAD | 130.1 s | 128.6 s | 116.2 s |
| DGX Spark (128 GB unified), NVFP4 QAD | 192.0 s | 198.3 s | 194.8 s |
| RTX 3090 (24 GB), FP8 | 16.9 min | 18.0 min | 14.5 min |
| Apple M5 Max (64 GB), INT6, MLX | 19.5 min | 20.1 min | 14.4 min |
A 5-second 832×480 clip with audio, end to end, with the 36-layer decoder. Spark times are three-seed means. Mac times are single runs. RTX 3090 times do not include about 7 minutes of model load per request. The 2× Spark runs used an earlier QAD checkpoint. Source: Hao AI Lab, "FastH3 Omni Goes Local", October 21, 2026; read October 23, 2026.
- Memory: DGX Spark keeps the full model in its 128 GB with no offload. The RTX 5090 keeps the NVFP4 transformer (18.5 GB) on the GPU and loads the text encoder only for the prompt. The RTX 3090 has no FP4 or FP8 hardware, so it runs FP8 weights with BF16 math and streams the transformer. The M5 Max peaks at 27–31 GB.
- Large reference sets: 3 images or a long reference clip exceed the 24–32 GB GPU profiles. One DGX Spark runs them at 480p in 640.3 s (3 images) and 504.6 s (1 long clip). The M5 Max takes 44–59 minutes.
- References cost time: one reference photo makes a request 3.3–4.4× slower than text to video on the same machine. On the RTX 5090, the time goes from 14.07 s to 61.9 s. Each reference image adds about 14,200 tokens to the 14,400 tokens of a 5-second 480p clip.
- Faster option:
--reference-max-pixels matchshrinks each reference to the clip size. On the M5 Max, the 1-photo request fell from 19.5 to 4.7 minutes. It changes framing and action, so it is off by default. - Serving: vLLM-Omni added the FastH3 series, including FastH3-Omni, on October 21, 2026 (PR #9167). Its single-card guide serves NVFP4 QAD weights on an RTX 5090, DGX Spark or RTX PRO 6000 Blackwell, and BF16 weights with online FP8 on an RTX 4090. An eight-GPU recipe serves the BF16 release. Both expose an HTTP endpoint that takes a prompt and a reference image and returns an MP4 with stereo audio (read October 25, 2026). SGLang added FastH3-Omni on October 22, 2026 (PR #45530). Its cookbook serves it on 2× H200 with CPU offload. A warm 5-second clip from one image takes 26.9 s at 480p and 46.6 s at 768p (read October 27, 2026).
Mac and multi-GPU
Mac, AMD and multi-GPU support
- Mac: ComfyUI did not ship Apple Silicon support at launch. FastVideo runs FastH3 V2 and Trim on Apple Silicon in MLX with INT6 weights (Hao AI Lab, October 6, 2026; read October 17, 2026). The only measured time is for FastH3 Trim: a 5-second 480p clip took 925.2 seconds on an M4 Max with 36 GB of unified memory. FastVideo gives no V2 time for the Mac.
- DGX Spark: FastH3 V2 took 141.4 seconds for a 5-second 480p clip on one Spark, and 87.2 seconds on two Sparks (same Hao AI Lab post). FastH3-QAD-v2 (October 19, 2026) took 58.1 seconds on one Spark and 33.6 seconds on two, with the 36-layer decoder (Hao AI Lab, October 21, 2026; read October 23, 2026). DGX Spark and RTX: FastVideo announced support for DGX Spark and RTX GPUs in the open FastH3-Omni release (announcement, October 13, 2026). The open weights and training code shipped on October 20, 2026 (FastVideo release post). FastH3-Omni local profiles, from the official model cards (BF16 release, NVFP4 QAD, FP8 for ComfyUI; read October 21, 2026): DGX Spark (128 GB unified memory) and RTX 5090 (32 GB) run the NVFP4 QAD weights in FastVideo, and both profiles default to 480p. The RTX 5090 offloads the text encoder and needs a precomputed AdaLN table. GPUs without FP4, such as the RTX 4090, use the FP8 recipe, and the bundled NVFP4 text encoder needs Blackwell. Measured FastH3-Omni times are in the next section. FastH3-Omni.
- Multi-GPU: SGLang and vLLM-Omni split H3 across GPUs. The GPU count must divide the 56 attention heads: 1, 2, 4, 7 or 8.
- Open weights output: a 768 px short side, 4–15 seconds, 24 fps. 2K output comes from H3-Regenerate-2K, which is API-only (MiniMax README).
Local vs dedicated host
When a dedicated host is the better choice
| Your own GPU | Nuva Lab dedicated host | |
|---|---|---|
| Best for | Tests and personal projects | Production volume for a team |
| Speed | Minutes per clip on one consumer GPU | About 4,000 ten-second 768p clips per day on 8× RTX PRO 6000 |
| Concurrency | One clip at a time | Every GPU serves your jobs |
| Customization | You run LoRA training yourself | We fine-tune and post-train on your data |
| Improvement | Manual | Data flywheel from your reviews |
| US companies | The H3 license excludes the US by default | MiniMax authorization for US companies |
Dedicated MiniMax H3 deployment · Cost per second at full use.
FAQ
Questions
How much VRAM does MiniMax H3 need?
MiniMax states no minimum. Full precision needs four 80 GB GPUs with SGLang. FastH3 and quantized builds run on 12–32 GB cards with offload, and they need a lot of system RAM. FastH3 Trim fits in 8 GB at 480p.
Can I run MiniMax H3 on an RTX 4090?
Yes, with quantized weights. FastH3 V2 in FP8 renders a 5-second 1344×768 clip in about 155 seconds on an RTX 4090, as measured by Hao AI Lab.
Can MiniMax H3 run on a Mac?
Yes, through the FastVideo MLX port with INT6 weights. FastH3 Trim took about 925 seconds for a 5-second 480p clip on an M4 Max with 36 GB of unified memory.
Can I run MiniMax H3 on 8 GB, 12 GB or 16 GB of VRAM?
Hao AI Lab measured FastH3 on one RTX 4090 with capped GPU memory. FastH3 V2 took 153.7 seconds at 16 GB and 170.7 seconds at 12 GB for a 5-second 1344×768 clip. FastH3 Trim fits in 8 GB and took 82.0 seconds for a 5-second 832×480 clip. A real card with less memory will be slower.
Is an RTX 5090 enough for MiniMax H3?
Yes, with FastH3 V2 in NVFP4. Hao AI Lab measured 14.8 seconds for a 5-second 832×480 clip and 38.6 seconds at 1344×768 on an RTX 5090 with 32 GB.
Can I run FastH3-Omni locally?
Yes. Hao AI Lab measured a 5-second 480p clip from one reference photo: 61.9 s on an RTX 5090 with NVFP4 QAD weights, 192 s on one DGX Spark, 16.9 minutes on an RTX 3090 with FP8 weights (plus model load), and 19.5 minutes on an Apple M5 Max (October 21, 2026). Requests with 3 images or a long reference clip need more than the 24–32 GB GPU profiles. DGX Spark runs them.
Can a US company run MiniMax H3?
The MiniMax H3 Community License lists the US as an excluded territory, so US companies need authorization from MiniMax. Nuva Lab holds a MiniMax authorization for H3 deployments that serve US companies.
Sources and further reading
Read next
Skip the hardware work
Tell us your volume. We'll run MiniMax H3 on a private host for your team.
Get started · 1 min