Run MiniMax H3 Locally on an RTX 3060: Tested, Six Traps

Short version: yes, MiniMax H3 runs on a 12GB RTX 3060, and the surprise is that VRAM is not the bottleneck. Peak VRAM was 4.7GB out of 12GB. System RAM was the constraint the whole way.

The cost is time. A 5-second 832x480 clip took 25 minutes 24 seconds. That makes local H3 excellent for unlimited free experimentation and impractical for producing finished work.

Everything below is from an actual run on one machine: RTX 3060 12GB, i5-12490F, 16GB DDR4, a DRAM-less Kingston NVMe, Windows with WSL2.

Read This First: The Licence Excludes the US, UK, EU and South Korea

Before you download 35GB, check whether you are allowed to use it. H3 ships under the MiniMax H3 Community License Agreement, not Apache or MIT. Clause 5 defines:

"Excluded Territories" means the European Union, the United Kingdom, the Republic of Korea and the United States of America.

And clause IV.4 goes further than most people expect — it covers not just the weights but what you make with them:

You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.

So in those four regions, the generated videos themselves are covered, not only the model files. MiniMax explains the reasoning in <a href="https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/QA-about-License.md" rel="nofollow" target="_blank">a Q&A document</a>: the EU AI Act is in force, the UK and South Korea are regulatory unknowns, and in the US MiniMax is in ongoing copyright litigation specifically about generative video. Their wording is "not yet, not never," and they invite people in those regions to contact them for a separate licence.

The hosted API is not restricted — it is globally available, because MiniMax runs that infrastructure and can enforce safeguards on it. If you are in an excluded territory, the API route below is the one open to you. More detail in our breakdown of what actually shipped with the open weights.

The Stack That Worked

Not native ComfyUI. We used <a href="https://github.com/deepbeepmeep/Wan2GP" rel="nofollow" target="_blank">WanGP (Wan2GP) v12.41+</a>, which is built for exactly this class of hardware, with Python 3.11.5 in a uv-created venv and torch 2.10.0+cu130.

The torch version is not optional — see trap 1. WanGP's own guidance for RTX 30-series is Python 3.11.x with PyTorch 2.10 and CUDA 13.0, and its docs explicitly warn off 2.8.0 (memory leak when switching models) and 2.9.0 (a 3D convolution issue that blows up VAE VRAM).

The minimum working model set is about 35GB, all from the community <a href="https://huggingface.co/DeepBeepMeep/MiniMax-H3" rel="nofollow" target="_blank">DeepBeepMeep/MiniMax-H3</a> repo — not the official one:

File	Size	Role
MiniMax-H3-FL2VA-pruned_rank8_int8_convrot.safetensors	21GB	DiT, pruned to ~20B, int8
qwen3vl-32B-MiniMax-H3-Q2_K.gguf	8.0GB	Text encoder
MiniMax-H3-video_vae_fp16.safetensors	4.9GB	Video VAE
MiniMax-H3-audio_vae_fp32.safetensors	578MB	Audio VAE

For scale: the official bf16 DiT is 66.28GB (33B parameters) and the full MiniMaxAI/MiniMax-H3 repo is roughly 498GB. ComfyUI's official bundle is 57GB. The community pruning and quantisation is what makes this fit at all.

Launch flags, all three required:

--profile 5 --perc-reserved-mem-max 0.1 --attention sdpa

Six Traps

These were each found by hitting the error, not by reading docs.

1. torch must be cu130, not cu128. On cu128 the int8 kernels silently fall back to eager execution. Community measurements put sampling at 6.12 s/it instead of 1.92 s/it — a 3.19× penalty for a mistake that produces no error message. Confirm it worked by looking for this line at startup:

[Quanto][INT8] Injected int8 kernels ACTIVE (backend=triton)

2. The default profile 4 goes straight to CUDA OOM. You need --profile 5. WanGP describes profile 5 as "VerylowRAM_LowVRAM (Fail safe): at least 24 GB of RAM and 10 GB of VRAM" — so 16GB of RAM is below the stated floor. It runs anyway, on the back of NVMe page file. That is also why a DRAM-less SSD hurts here.

3. SageAttention fails outright on H3 for RTX 30-series. Sage 1.0.6 throws AssertionError: varlen only support head_dim [64, 128]. RTX 30xx cannot use Sage 2, so the only option is --attention sdpa. On 40- and 50-series, Sage 2.2 works and community reports put the saving around 30%.

4. hf_xet stalls large downloads at zero bytes. The video VAE sat at 0 for eight minutes. Setting HF_HUB_DISABLE_XET=1 restored it to 18MB/s immediately.

5. GGUF llama.cpp CUDA kernels are unavailable — the log says [GGUF][llama.cpp CUDA] kernels unavailable, using fallback. So use the int8_convrot DiT rather than a GGUF DiT. GGUF is still the right call for the text encoder, where Q2_K saves a lot of RAM.

6. A system proxy will intercept localhost and Gradio dies with a 503: Couldn't start the app because 'http://localhost:7860/gradio_api/startup-events' failed. Set NO_PROXY=localhost,127.0.0.1; outbound traffic still goes through the proxy.

Bonus, if your C drive is tight: redirect UV_CACHE_DIR, TMP, TEMP and HF_HOME to another drive before you start.

Measured Performance

All on the same 3060, 20 steps:

Resolution / frames	Time per clip	Per step
608x352 / 73 frames (3.0s)	19 min 50 s	—
832x480 / 124 frames (5.2s)	25 min 24 s	71.8 s
864x480 / 124 frames (5.2s)	28 min 38 s	84 s
1280x736 / 192 frames (8s)	~3 h (estimated, not run)	—

Peak VRAM: 4.7–4.8GB of 12GB. Free system RAM: 0.2GB. That is the whole story — the GPU is idling while the machine swaps.

The first row is the one worth staring at. Dropping to 608x352 with 73 frames cuts the token count to roughly 30% of the 864x480 run, and yet the time only falls from 28.6 minutes to 19.8 — nowhere near proportional. The reason is that on a 16GB machine every step re-streams the 21GB of weights from NVMe, and that overhead is fixed regardless of resolution. So "just render smaller" barely helps when you are RAM-limited. Adding RAM is the fix; shrinking the canvas is not.

The comparison that proves it: the ComfyUI team measured 480p/5s at roughly 9 minutes on the same 3060 with 32GB of RAM. Doubling system memory is worth about 2.8× here. If you are choosing what to upgrade, buy RAM, not a bigger GPU.

Output verified with ffprobe: h264 video plus AAC stereo at 32kHz — genuine native audio generated in the same pass, not added afterwards — 24fps, 124 frames = 5.167 seconds.

Constraints You Cannot Prompt Around

• No 2K locally. The open weights render on a short-side-768 canvas, maximum 768x1344. The advertised 2K comes from an API-side In-Context Regeneration pass.
• Frame counts must be 17k+5: 73, 124, 175, 192.
• Both dimensions must be multiples of 32. This is why 720p is written 1280x736 — 720 is not divisible by 32.
• Duration scales badly. Community numbers from a 5090: 5s = 168s, 7s = 282s, while 15s runs anywhere from 7 to 50 minutes depending on tuning. Three 8-second clips are far cheaper than one 15-second clip.
• No distilled or lightning LoRA yet as of early August 2026, so 15–20 steps is the floor. Quantised builds are appearing quickly though — GGUF Q4_K_M at 19.86GB, Q3_K_M at 15.57GB, plus NVFP4 and INT4 variants.

Other hardware, from community reports: a 5090 with SageAttention does 1280x736/5s in 168s (242s without), and with SM120 sparse attention plus EasyCache, 15 seconds drops from 55 minutes to about 7. A tuned 4090 does 362 frames in 9.6 minutes. An RTX PRO 5000 peaks at 46.5GB with no offloading at all.

Local, Rented, or API?

For an 8-second clip:

Route	Cost per clip	100 clips
MiniMax API, 768p ($0.09/s)	$0.72	$72
MiniMax API, 2K ($0.13/s)	$1.04	$104
Rented 5090 (~$0.21–0.54/h, ~5.5 min/clip)	~$0.023	~$2–5
Local 3060	electricity only	~3 h each

Renting is 15× to 50× cheaper than the API depending on resolution and card — a gap large enough to reorganise a workflow around. The sensible split: run locally to iterate, because failed takes cost nothing and prompt refinement is mostly failed takes; rent a 5090 to render finals; use the API when you need 2K, or when the licence means the weights are not an option for you.

Get Your Prompts Right Before You Spend 25 Minutes

At 25 minutes a take, prompt quality matters more locally than anywhere else. The structures that work on H3 — explicit camera sequencing, a named closing frame, sound described as deliberately as picture — are broken down in our H3 launch prompt analysis.

You can also browse what creators are actually running, sorted by model, on trending prompts.

And if you have a clip you want to reverse-engineer rather than guess at, that is what our video to prompt tool is for: paste a link or upload the file and it returns a structured prompt covering scene, motion and camera work — which you can then feed to H3 locally, for free, as many times as you like.