MiniMax H3 Hardware Map: What Runs, What Doesn't, and Why the Answer Is Probably Your Card
MiniMax H3 is a 33B video model with native stereo audio. This operator's hardware map covers what runs it, what doesn't, the territory restriction, and the right rig for you.
MiniMax H3 Hardware Map: What Runs, What Doesn’t, and Why the Answer Is Probably Your Card
Audience: Operators running or planning to run MiniMax H3 (33B video model with native stereo audio) locally, or anyone evaluating H3 vs cloud video generation.
Date: 2026-08-08 · Status: draft · Last verified against primary sources: 2026-08-08
A note on this snapshot. Prices, software support, license territory, and benchmark numbers all change. Before you cite anything from this article, check the dossier at
Websites/content-strategy/research/minimax-h3/minimax-h3-research-2026-08-08.md. The dossier is the source of truth. This article is the readable form.Most important caveat — territory restriction. MiniMax H3 open weights are excluded for use in the United States, European Union, United Kingdom, and South Korea per the HuggingFace model card license. Operators in those territories must use the cloud API path. Local installation is not legally available to them. This is the single most important operator finding in this article. If you’re in an excluded territory, jump to the “Cloud API” section.
The one-paragraph version
MiniMax H3 is a 33B-parameter video model released July 31, 2026, that generates 5–15 second clips at up to 2K with native stereo audio in the same pass — not bolted on afterward. It’s the first MiniMax release with open weights and the first frontier video model that does audio + video in one generation. The hardware map by Yume_X (Aug 7, 2026) measured every consumer GPU and found that the floor is a $300 RTX 3060 12GB (2 hours for a 10-second clip) and the ceiling is the 5090 (2.5 minutes for 7 seconds). The binding constraint for H3 is not VRAM alone — it’s VRAM × system RAM × residency strategy. This guide is the operator’s view: what runs, what doesn’t, what the territory restriction means, and what the right answer is for you.
What H3 actually is
The model
H3 is MiniMax’s third-generation video model and the first they’ve released with open weights. Key specs from the MiniMax blog and Comfy.org Day-0 announcement:
| Spec | Value |
|---|---|
| Parameters | 33B |
| Modalities | Text → video, image → video, video → video, audio + video (combined) |
| Output resolution | Up to 2K |
| Output duration | 5–15 seconds (configurable) |
| Audio | Native stereo, generated in the same pass as video |
| Reference inputs | Up to 9 images + 3 videos + 3 audio files (12-file cap) |
| Multi-shot | Native — one prompt can composite multiple shots |
| In-context editing | Edit character/object/background/dialogue within an existing clip |
| Architecture innovations | Contextual Omni Representation · H3-VAE (48-channel latent) · H3-Omni Transformer |
What the audio capability means
H3 generates audio in the same pass as video. Speech, music, ambient sound, lip sync, voice timbre transfer — all in one generation. This is the unique value H3 has over competitors.
If your project needs video + audio in one shot (no post-process assembly, no separate audio generation, no timeline sync work), H3 has no current cloud alternative cheaper at 2K. If you can stitch audio after the fact, competitors are competitive on price.
What it costs
| Provider | Price | Per 5s clip | Per 15s clip |
|---|---|---|---|
| Hailuo API (our standard) | $0.10/clip | $0.10 | $0.30 |
| MiniMax H3 2K direct | $0.13/sec | $0.65 | $1.95 |
| fal.ai H3 | $0.26/sec | $1.30 | $3.90 |
For 5 Tier-1 clips at 2K with audio: Hailuo API $0.50 vs MiniMax H3 direct $3.25 (6.5× more expensive) vs fal.ai $6.50 (13× more expensive). For low-volume Tier 1 work, Hailuo is the right answer.
Note: the MiniMax Token Plan does NOT cover H3 (Error 2013). The Token Plan documentation says it covers all models, but H3 is excluded. Pay-per-use direct API is the only Token Plan path.
The territory restriction — read this first
Per the HuggingFace model card license for MiniMaxAI/MiniMax-H3:
Excluded territories: United States, European Union, United Kingdom, South Korea.
For operators in those territories, the local installation path is closed. The cloud API is available globally. If you’re in an excluded territory, your only legal options are:
- MiniMax H3 API direct ($0.13/sec for 2K)
- fal.ai H3 ($0.26/sec)
- Hailuo API ($0.10/clip, our standard Tier 1 path)
The Yume_X hardware map does not flag this. If you read the Yume post and assumed the local path is available globally, that assumption is wrong.
The hardware map (12 GPU paths, real measured timings)
This is the Yume_X map reproduced with one correction and one addition.
| GPU | Memory | Measured H3 behavior |
|---|---|---|
| RTX 5060 8GB | 8GB | Floor. Full 15s at 480p · 10s at 540p · 5s at 720p. 32GB system RAM alongside. |
| RTX 3060 12GB | 12GB | Old floor. 10s generation in 2 hours. It finished. |
| RTX 4070 12GB | 12GB | 608×352, 20 steps, 167 seconds. Low res but functional. |
| RTX 4090 Laptop 16GB | 16GB | 960×540, 5s, 182 seconds with SageAttention. |
| RTX 5060 Ti 16GB | 16GB | 323s baseline for 5s @ 864×480. SageAttention: 187s. SageAttention + EasyCache: 41s (skips 7 of 20 steps — quality cost). |
| RTX 5070 Ti 16GB | 16GB | The --fast-disk flag cut system RAM usage 3.6×. Peak VRAM ~14.4 GiB for 5s @ 1344×768. |
| RTX 3090 24GB | 24GB | Full 15s clip + audio on a single 3090 with only 31GB system RAM. 23 minutes 17 seconds for the full render. The fix was one flag: --disable-pinned-memory drops RAM from 29.8GB to 7.5GB. |
| RTX 4090 24GB | 24GB | ~100 seconds warm T2V. 12s HD clips 10–13 minutes with SageAttention. |
| RTX 5090 32GB | 32GB | Consumer sweet spot (outside excluded territories). NVFP4 pack: 864×480, 10s, 175s, peak 26.9 GiB. ~2.5 minutes for 7s T2V. ~7 minutes for 5s I2V. 500W power limit faster than 600W (thermal throttling) — 17% less energy. |
| 4× H100 | 80GB+ | 13 seconds. |
| 8× B300 | 80GB+ | 18 seconds. |
| 2× DGX Spark (128GB each) | 128GB unified | Single Spark: ~155s. Dual Sparks: 64.9–68.8s (2.3× speedup). Both GPUs at ~96% during denoise. Ulysses sequence parallelism over RoCEv2, 100W total power. |
Yume’s correction we should remember: the binding constraint is VRAM × system RAM × residency strategy, not VRAM alone. Three receipts:
- The 3090 died on 31GB of system RAM, not on 24GB of VRAM. Peak VRAM was 19.8GB. ComfyUI’s 90% system RAM pin caused the SIGKILL.
- The 5070 Ti flag moved weights into reclaimable page cache and cut RAM 3.6×. The
--fast-diskflag is the lever. - ComfyUI’s pruning cut the model footprint 66%, from 123.6 GB to 42.5 GB. Every “how to run it” lane is a RAM residency strategy wearing a GPU label.
The Blackwell compounding stack
On Blackwell cards (5090, 5080, 5060 Ti), the acceleration stack compounds:
| Lever | Speedup | Quality cost | Requirement |
|---|---|---|---|
| SageAttention | 30–40% | None | int8 kernel hits CUDA fast path |
| EasyCache | 2× | Skips denoise steps — quality drops on complex motion | Optional |
| INT8 ConvRot | ~10% faster than FP8 | None (counter to expectation) | CUDA 13.0+ required |
| Power limit reduction | 17% less energy at 500W vs 600W | None (thermal throttle at 600W) | BIOS or driver setting |
| Step count reduction | 50→10 steps cuts time 47% | None for typical outputs | Default change |
The cheapest lever is step count. Default consumer-hardware step count should be 10–20, not the 50 used in official server recipes. Nobody on consumer hardware should use 50 steps.
The three setup paths (with foot-guns)
Path A — ComfyUI + NVIDIA (NVIDIA owners)
When to use this path: You have a CUDA-capable NVIDIA GPU with ≥8GB VRAM, you’re outside the MiniMax-excluded territories, and you want maximum control over the generation.
Weights needed (5 files, ~39.55 GB total):
| File | Size | Purpose |
|---|---|---|
minimax_h3_fl2va_pruned_int8_convrot.safetensors | 19.5 GB | Main H3 transformer |
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | 14.6 GB | Text encoder |
minimax_h3_video_vae_fp16.safetensors | 4.85 GB | Video VAE |
minimax_h3_audio_vae_fp32.safetensors | 0.56 GB | Audio VAE |
ref2va.safetensors | 19.5 GB | Only needed for R2V (reference-to-video) mode |
Setup steps:
- Install ComfyUI 0.30.0+ (ships with
MiniMaxH3ReferenceToVideoandMiniMaxH3I2Vcustom nodes). - Download the weight files to your ComfyUI models directory.
- Load the official H3 workflow JSON.
- Set step count to 10–20 (not 50).
- Generate.
Foot-guns:
- Comfy-Org#15315 (still open as of Aug 8, 2026): H3 T2V on Apple Silicon MPS produces black tiles or completely black video. Do not use ComfyUI on Apple Silicon for H3.
- ComfyUI 90% system RAM pin: Causes silent SIGKILL on rigs with limited system RAM. Use
--disable-pinned-memoryif you have <32GB system RAM. - EasyCache + complex motion: EasyCache doubles speed but skips denoise steps. Quality drops on complex motion. Use only when motion is simple.
- Reference mode + endpoint-frame mode: Mutually exclusive. Pick one.
- Reference + duration controls: Cannot both be set. Pick one.
Path B — Sawfwair MLX + mere.run (Apple Silicon only, beta)
When to use this path: You have an Apple Silicon Mac (M1 Max / M2 Ultra / M3 Ultra / M4 Max / M4 Pro), you want to run H3 locally, and you’re willing to be a beta tester.
Weights: sawfwair/MiniMax-H3-FL2VA-MLX-4bit — 43 GB total, including text encoder (8-bit), transformer (4-bit), video VAE, audio VAE, AdaLN cache.
Validated hardware: Mac Studio M4 Max 128GB (peak 27.7 GB RAM resident, 18:54 for 2.3s clip at 512×512, 22 dB PSNR).
Unvalidated hardware: Mac mini M4 Pro 48GB. Our install was successful (all 14 SHA256 hashes verified), but our attempts to generate a clip returned “Orphan recovery” three times in a row, and we experienced one Mac freeze during model loading. We never produced a successful clip on M4 Pro 48GB.
Setup:
# 1. Build the mere.run Swift CLI
git clone https://github.com/sawfwair/mere-run
cd mere-run && swift build -c release
# ~6 minutes, 104.5 MB binary
# 2. Download the Sawfwair 4-bit weights
# (all 14 SHA256 hashes should verify on first download;
# re-download failed files if any mismatch)
# 3. Generate
mere.run video generate "your prompt here" \
--model-root /Users/youruser/Models/Sawfwair-H3-MLX-4bit \
--width 512 --height 512 --num-frames 56 --fps 24 \
--seed 314159 --steps 31 --h3-weight-mode quantized \
--output /path/to/output.mp4Foot-guns (these are real, we hit them):
- Orphan recovery: the process never stays alive past the parent process lifecycle. Use
nohuporscreen, don’t run as foreground process. - Mac freeze during model load: 28GB of weights loading into RAM can freeze the system. Force inactive eviction first (
purgeor allocate-then-free). - Competing services:
mlx_lm.server(1.1 GB) andvoicebox-server(376 MB) compete for RAM. Stop them before running mere.run. - Note:
mere.runis a general local-AI toolkit, not a MiniMax CLI. Our earlier framing of it as “the MiniMax CLI” was inaccurate.
Path C — Cloud API (excluded-territory operators + low-volume users)
When to use this path: You’re in a MiniMax-excluded territory (US, EU, UK, S. Korea), or you need a video today with no setup, or you’re producing fewer than 100 clips per month.
Providers (Aug 2026 pricing):
| Provider | Per-second | Per 5s clip | Per 15s clip | API key |
|---|---|---|---|---|
| Hailuo API | n/a (per clip) | $0.10 | $0.30 | Hailuo platform |
| MiniMax H3 2K direct | $0.13/sec | $0.65 | $1.95 | platform.minimax.io |
| fal.ai H3 | $0.26/sec | $1.30 | $3.90 | fal.ai |
Setup: Get an API key from the provider. Make a POST request with your prompt, your width/height/frames, and the model name. Receive the MP4 in 30–90 seconds.
Foot-gun: MiniMax Token Plan does NOT cover H3 (Error 2013). Pay-per-use direct API only.
The one-paragraph decision tree
Start with your jurisdiction. If you’re in a MiniMax-excluded territory, your answer is the cloud API — Path C. If you’re outside the excluded territories, look at the GPU you already own and check if it’s on Yume’s map. If it is, run Path A (NVIDIA) or Path B (Apple Silicon). If it isn’t, buy the cheapest GPU from Yume’s map that fits your budget: 5060 8GB at $300 for the lowest entry, 5090 32GB at $2,000 for the sweet spot, or 4× H100 at hyperscaler prices for the server lane.
If you only need 5 clips a month, Path C at $0.10/clip is the cheapest path. The “local is cheaper” intuition is wrong for low-volume users.
What the right answer looks like for each persona
Hobbyist with a $300 budget
Get an RTX 3060 12GB. Run H3 via ComfyUI at 10s generations in 2 hours. Accept that this is slow. Use it as a learning platform, not a production tool.
Indie developer with a $2,000 budget
Get an RTX 5090 32GB. Use SageAttention, use step count 10–20, use the NVFP4 pack. You’ll get 7s T2V in ~2.5 minutes. Set power limit to 500W.
Small-shop operator with $4,000
Get a DGX Spark 128GB. Full CUDA, 273 GB/s bandwidth, 100W power. 70B NVFP4 at 5.39 tok/s for LLM inference; H3 video at the Yume-measured ~155s on single Spark. Best balance of capability and power efficiency for a serious local setup.
24/7 inference operator
128GB Mac Studio M2/M3 Ultra. 50W peak, $280/year electricity. 12.5 tok/s on 70B Q4 for LLM inference; H3 video at Sawfwair-MLX 4-bit on Apple Silicon is beta but the cheapest 24/7 path for Apple-Silicon-native shops. Plan for cloud-API fallback if Sawfwair MLX proves unstable in your specific setup.
Operator in excluded territory (US, EU, UK, S. Korea)
Cloud API. Hailuo at $0.10/clip for Tier 1 standard work; MiniMax H3 direct at $0.65/clip when you specifically need 2K + native audio; fal.ai at $1.30/clip when MiniMax direct is rate-limited.
What H3 is bad at
For honesty, these are the things H3 doesn’t do well. The Trilogy AI review documented them. Knowing what H3 can’t do matters more than knowing what it can.
- Anatomy and biology. Tested with “impossible heart” prompt — speech was accurate, heart anatomy was so bad the clip had to be discarded. The model’s surface polish hides the failure.
- Geometry and exact shapes. Paper boat shape broke inside the clip. Cicada emergence looked plausible but not entomologically accurate.
- Long-form narrative. H3 is best for 5–15 second clips. Anything longer requires careful prompt engineering or multiple clips stitched together.
- Character consistency in long sequences. Identity drift in complex scenes with multiple characters + rapid camera movement.
- Resource intensity. 42.5 GB peak VRAM at smallest config. Dynamic VRAM offloading helps but is slow on consumer hardware.
If any of these matter for your project, H3 is the wrong tool. Use a competitor.
What this guide does not cover
- The Yume_X post is the canonical reference for H3 hardware measurements. All 12 GPU timings in this article come from that post or from community benchmarks cited in it. Read it directly: https://x.com/yume_arasaki/status/2085802508526789060
- Training. This is inference-only.
- ComfyUI workflow customization. The official MiniMax H3 workflow JSON is the starting point; customization is a craft.
- Voice cloning or face replacement. H3 supports voice timbre transfer but face cloning is outside scope.
- Music video production. H3 generates audio but the audio is generated to match the video, not the reverse.
Sources
Primary operator reference:
- Yume_X (@yume_arasaki), “MiniMax H3 Hardware Map” X thread, Aug 7, 2026 — https://x.com/yume_arasaki/status/2085802508526789060
- Yume_X “MiniMax H3 Hardware Map — Sources and Recipes” GitHub Gist, Aug 8, 2026 — gist.github.com/yume-arasaki
H3 model + license:
- MiniMax H3 official blog — https://www.minimax.io/blog/minimax-h3
- HuggingFace model card (license territory restriction) — huggingface.co/MiniMaxAI/MiniMax-H3
- Comfy.org Day-0 announcement — blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui
HuggingFace model repos:
- MiniMaxAI/MiniMax-H3 (official) — huggingface.co/MiniMaxAI/MiniMax-H3
- sawfwair/MiniMax-H3-FL2VA-MLX-4bit (Apple Silicon 4-bit MLX port) — huggingface.co/sawfwair/MiniMax-H3-FL2VA-MLX-4bit
- pipenetwork/MiniMax-H3-MLX-8bit (Apple Silicon 8-bit alternative) — huggingface.co/pipenetwork/MiniMax-H3-MLX-8bit
ComfyUI + Apple Silicon bugs:
- Comfy-Org/ComfyUI#15315 — H3 T2V on Apple Silicon MPS broken (still open Aug 8, 2026) — github.com/Comfy-Org/ComfyUI/issues/15315
- Comfy-Org/MiniMax-H3 discussion #24 — dequantize_int8_embedding bug (different from #15315, fixed) — github.com/Comfy-Org/MiniMax-H3/discussions/24
Community reviews:
- Trilogy AI Substack (Leonardo Gonzalez, Aug 1 2026) — trilogyai.substack.com/p/minimax-h3-ai-video
- PixVerse review (Jul 30 2026) — pixverse.ai/en/blog/minimax-h3-review
- AI Inspo review — ai-inspo.com/blogs/model/minimax-h3-review
Pricing references (Aug 2026):
- Hailuo API pricing — platform.minimax.io
- MiniMax H3 2K direct pricing — platform.minimax.io
- fal.ai H3 pricing — fal.ai/learn/devs
- Note: MiniMax Token Plan does NOT cover H3 (Error 2013) — platform docs
Inference backends:
- llama.cpp — github.com/ggml-org/llama.cpp
- Apple MLX framework — github.com/ml-explore/mlx
- mere.run CLI — github.com/sawfwair/mere-run
- MLX arXiv paper (primary source) — arxiv.org/abs/2511.05502
Our operator experience (Day-1 attempt log):
Companion dossier + verification files:
Websites/content-strategy/research/minimax-h3/minimax-h3-research-2026-08-08.md— full evidence ledgerverify-model-spec.md— HF repos, license territory, weight filesverify-hardware-paths.md— 12-row hardware × axes matrixverify-setup-paths.md— 3 paths + first-run commandsverify-promotion-strategy.md— Yume-vs-ours mapping, cost-amortization- Companion guide: “How I ran H3 on a 48 GB Mac mini (and how you can on whatever you have)” — the operator-journey follow-up article, available at the same R2 prefix
Honest disclosure
This article is built on the Yume_X hardware map. Yume’s map gets 11 of 12 things right. The one thing it does not flag is the territory restriction. If you read the Yume post and assumed the local path is available globally, that assumption is wrong. For US/EU/UK/S. Korea operators, the local installation is not legally available; the cloud API is the only path.
Our own operator experience (a 48 GB M4 Pro Mac mini) did not produce a successful local H3 clip. We hit orphan-recovery process issues and a Mac freeze during model loading. We pivoted to the Hailuo API for our Tier 1 video work. The honest answer for our specific use case (5 clips, time-pressed, low budget) was the cloud API. The honest answer for a different operator (Apple Silicon-native shop, 24/7 inference, willing to debug beta software) might be the Sawfwair MLX path.
This article’s job is to give you the full landscape so you can pick the path that fits your situation. Refresh the dossier before citing any of this in published copy. Prices and software support change monthly.
Last verified: 2026-08-08. Refresh quarterly or before any published copy cites it.
Related guides
These guides extend the work above:
Sources
Submit a take
Have a different read on this? Drop a comment below — your email isn't published, and I read every one. Nothing leaves the site until I approve it.