MiniMax H3 is one of the most convincing attempts yet to turn AI video generation into a single multimodal production system. It accepts text, images, video and audio as context, then generates up to 15 seconds of video with native 32 kHz stereo sound. It can animate a first frame, bridge a first and last frame, or use as many as 12 mixed reference files to guide subjects, motion, voice and sound.
The downloadable release is also much narrower than the launch pitch first suggests. MiniMax has released H3-Base, which produces video at a 768-pixel short edge. The company’s hosted H3-Context-IR system and H3-Regenerate-2K stage are not in the weight release. The public full-2K recipe therefore sends work back to MiniMax’s API. Running the original BF16 base model locally also means accommodating roughly 134 GiB of weights for one task partition before activation memory and runtime overhead.
Early independent evidence is strong. H3 entered the top three in all six relevant Artificial Analysis preference pools and ranked first for audio-inclusive video editing when checked on August 3. Those scores measure the hosted pipeline, however, not a local render from the downloadable base checkpoint or a community quantization. MiniMax has published no comparable model-quality benchmark table, multi-seed failure analysis, safety evaluation or technical report.
Testing disclosure: Kingy did not run MiniMax H3 or generate test video for this article. The analysis uses official files and documentation, framework-published performance measurements and an independent blinded preference arena, all checked August 3, 2026.
The answer in one sentence: MiniMax H3 is a frontier-competitive hosted audio-video generator and an unusually capable open-weight local base model, but buyers and developers must distinguish hosted H3 from local H3-Base, budget for heavy compute, and read a community license that excludes the US, EU, UK and South Korea.
Quick verdict
H3 is most interesting for teams that need video and synchronized sound from the same generation pass, especially when a shot must follow several kinds of reference material. It is less attractive as a casual local model. The official lossless consumer recipe needs two RTX 5090s and a 384 GiB-class host; smaller single-GPU routes depend on quantization, aggressive CPU offload and lower resolutions.
The hosted API is the sensible place to judge H3 before buying hardware. At the prices checked for this article, a 15-second generation costs $1.20 at 768P or $1.95 at 2K before billable video references, extra images or Context-IR tokens. That is cheap enough to validate the workflow and output quality with real production prompts.
| Best for | Skip or wait if |
|---|---|
| Multimodal shots that combine text, images, motion references and audio | You need a fully offline version of MiniMax’s advertised 2K pipeline |
| Native dialogue, ambience, effects or music generated with the video | You are deploying the weights in the US, EU, UK or South Korea without a separate licence |
| First-frame, last-frame and semantic reference workflows in one model family | You need frame-faithful video editing or exact source timing |
| Teams with datacenter GPUs or patient, high-RAM consumer offload setups | You expect a small, fast, Mac-ready or CPU-friendly local model |
| Pros | Cons |
|---|---|
| Strong early blinded human-preference results | No official comparative quality evals or technical report at launch |
| Video and native stereo audio generated together | Hosted arena scores do not validate the local base checkpoint |
| Broad text, image, video and audio conditioning | Context-IR and 2K regeneration are not open-weight |
| Official SGLang, vLLM, Diffusers and ComfyUI integration work | One original task partition is about 144 GB on disk |
| Useful first/last-frame and semantic-reference modes | Initial open release lacks MiniMax’s trained sparse-attention implementation |
| Hosted price is low enough for practical evaluation | Restrictive territory, commercial-display and model-training terms |
What exactly was released?
The dates require care. MiniMax published its launch article on July 31, 2026. The community licence calls August 2 the release and licence date. The official Hugging Face repository appeared publicly on August 3. The official materials do not reconcile the three dates.
The full product has three stages:
- H3-Context-IR interprets free-form combinations of text, images, video and audio. MiniMax describes it as a hosted, multi-stage system that parses instructions, associates content across modalities and turns the result into a structured intermediate prompt.
- H3-Base generates video and stereo audio at a 768-pixel short edge. This is the part MiniMax released as weights.
- H3-Regenerate-2K feeds the 768p result and the original context back through H3 to regenerate a higher-resolution output. This stage remains hosted.
That separation changes the meaning of “open.” H3 is an open-weight base model under a custom licence, not a fully open implementation of the commercial system and not OSI open-source software. A local H3-Base render can be entirely local. The official full-2K workflow cannot.
MiniMax provides two separate base checkpoints. FL2VA handles text-to-video-and-audio plus first-frame, last-frame and first-and-last-frame generation. Ref2VA handles semantic reference generation from images, videos and audio. Video-to-video is a Ref2VA use case, not a separate public checkpoint.
Specifications at a glance
These are published specifications, not Kingy measurements. The API and local frameworks impose slightly different operational limits, so the table uses the most concrete official value and calls out the boundary.
| Specification | MiniMax H3 | Evidence and limit |
|---|---|---|
| Output duration | 4–15 seconds | Local frame counts align to the VAE’s 17n+5 grid; a nominal five-second SGLang run is 124 frames, or 5.167 seconds |
| Output frame rate | 24 fps | Canonical local and API output rate |
| Output audio | 32 kHz stereo | Video and audio latents are predicted jointly |
| Local resolution | 768-pixel short edge | Common 16:9 canvas is 1344×768; exact dimensions align to multiples of 32 |
| Hosted high resolution | “2K” | Produced through H3-Regenerate-2K; English docs do not define the label as one exact pixel dimension |
| Aspect ratios | Includes 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16 | API accepts a broad set rather than one fixed canvas |
| Dialogue languages | 11 described as stable | Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish; others vary |
| First/last-frame inputs | Zero, one or two images | FL2VA: text only, one endpoint image, or both endpoint images |
| Semantic references | Up to 12 mixed files | Up to 9 images, 3 videos and 3 audio clips; video and audio duration totals are each capped at 15 seconds |
| Reference audio | Up to three clips | Audio cannot be the only reference; it must accompany an image or video |
| Prompt limit | 7,000 characters per text item via API | The bundled encoder’s 262,144-position config is not an advertised H3 prompt window |
| Local checkpoint precision | BF16 with some FP32 components | Released as safetensors; quantized ComfyUI conversions are separate repacks |
| Generation steps | 50 in published SGLang benchmark recipes | The released checkpoints are CFG-distilled, so each step uses one forward pass and exposes no negative prompt or guidance scale in the Diffusers design |
Source: MiniMax H3 model card, H3 API specification and SGLang’s H3 guide, checked August 3, 2026.
Architecture: why the “33B model” needs far more than 66 GB
H3-Base packs text, reference media, target-video latents and target-audio latents into one multimodal sequence. A single dense transformer processes that sequence and jointly predicts the audio and video latents. This is more direct than stitching together a silent video model, a speech model, a Foley generator and a synchronizer after the fact.
The H3-Omni-Transformer has 33 billion parameters. It is a dense, single-stream, 50-layer model with a hidden size of 5,376, 56 attention heads and three-dimensional rotary position embeddings across time, height and width. Around 13 billion parameters sit in AdaLN-related branches. MiniMax says their outputs can be precomputed and cached for inference, which explains why pruned inference conversions can be much smaller than the complete training-capable transformer release.
That 33B figure excludes the model’s full Qwen3-VL-32B encoder and both VAEs. The encoder supplies hidden states from its 50th layer. The visual VAE uses 16× spatial and 4× temporal compression, then H3 patchifies its latents to an effective 32× spatial reduction before the transformer. The audio VAE compresses each 32 kHz stereo channel into a 40 Hz latent stream.
The result is an inference stack of roughly 69.2 billion parameters across the encoder, generator and VAEs. From the official safetensors headers and file sizes, one task partition breaks down as follows:
| Component | Stored weight size | Role |
|---|---|---|
| Qwen3-VL conditioner | 62.13 GiB | Interprets text and visual context |
| H3 Omni Transformer | 61.73 GiB | Joint audio-video denoising |
| Visual VAE | 9.70 GiB | Encodes and decodes video latents |
| Audio VAE | 0.56 GiB | Encodes and decodes stereo audio latents |
| Total | 134.13 GiB | Weight files only, before runtime overhead |
MiniMax trained H3 with sparse attention, but the initial weight release exposes full-attention inference only. That matters because video length and resolution grow the packed sequence quickly. For a 1344×768, 124-frame target, the visual stream alone is roughly 31,248 transformer tokens after temporal compression and patchification. A 15-second aligned output is around 91,000 target-video tokens before text, audio or references. Tripling the duration roughly triples the sequence length while the full-attention matrix can approach nine times the work.
H3 uses a diffusion transformer rather than autoregressive video-token generation, so there is no ordinary LLM KV cache. It repeatedly evaluates the transformer over a long multimodal sequence. FlashAttention-class kernels, sequence parallelism, careful placement and VAE tiling are central to making it run.
Benchmarks and evals: the independent signal is strong, the official evidence is thin
MiniMax did not publish a numerical model-quality benchmark table with H3. The launch article claims strong instruction following, text rendering, brand rendering, motion transfer and price-performance, but names no comparison set, sample count, scoring rubric or confidence interval. The model card provides reproducible example scripts and outputs, not an evaluation suite. A full technical report was still “coming soon” at publication.
The best independent evidence available on release day came from Artificial Analysis. Its arena shows two anonymous videos created from the same prompt and asks people to choose a preference. A Bradley–Terry model converts those pairwise votes into an Elo-like rating. The site separates text-to-video, image-to-video and video-editing pools, with additional pools in which voters also judge audio.
These were H3’s live results when retrieved on August 3, 2026. The leaderboard recalculates regularly, so the point estimates and ranks may move.
| Evaluation pool | H3 result | Nearest comparison | What the result supports |
|---|---|---|---|
| Text-to-video, audio judged | #2; 1239.36 Elo; 95% CI 1229.36–1249.36; 6,035 appearances | Gemini Omni Flash: 1244.07 | H3 is in the top preference cluster; overlapping intervals do not establish a meaningful gap |
| Text-to-video, audio not judged | #2; 1305.75; CI 1294.75–1316.75; 5,171 appearances | Gemini Omni Flash: 1324.23 | Frontier-level preference, with narrowly overlapping intervals |
| Image-to-video, audio judged | #3; 1186.52; CI 1176.52–1196.52; 5,383 appearances | Seedance 2.0 720p: 1195.77; Gemini Omni Flash: 1192.39 | The top three are statistically difficult to separate from these intervals |
| Image-to-video, audio not judged | #2; 1350.96; CI 1338.96–1362.96; 5,098 appearances | Gemini Omni Flash: 1368.40 | H3 sits between Gemini and Seedance, with uncertainty overlapping both |
| Video editing, audio judged | #1; 1130.25; CI 1124.25–1136.25; 8,214 appearances | Gemini Omni Flash: 1121.85 | H3 ranked first, but the confidence intervals still overlap |
| Video editing, audio not judged | #3; 1097.83; CI 1088.83–1106.83; 8,106 appearances | Gemini Omni Flash: 1137.33; Wan 2.7: 1106.16 | Behind Gemini’s point estimate and close to Wan within the uncertainty |
Sources: Artificial Analysis text-to-video, image-to-video and video-editing leaderboards.
An “appearance” is one side of a matchup, not a unique prompt or an independent laboratory trial. Audio and non-audio ratings belong to separate pools and cannot be compared numerically. Artificial Analysis also notes that Gemini refused some editing prompts, particularly those involving people; its editing score covers successful generations rather than the full submitted set.
The arena evaluates a hosted H3 endpoint. Its default methodology targets 10-second, 16:9 output near 1080p, uses creator-recommended prompt enhancement and accesses first-party APIs or repositories. H3’s hosted pipeline can use Context-IR and high-resolution regeneration that the weight release omits. The scores therefore support calling the hosted H3 product frontier-competitive. They do not prove that a local 768p H3-Base render, a pruned INT8 checkpoint or a different prompt processor will match it.
What has not been evaluated publicly
No public H3 evidence found for this review answers these questions:
- What are the VBench, VBench-2, VideoPhy or EvalCrafter scores under a disclosed protocol?
- How often do dialogue, lip synchronization, sound effects and music fail across languages and shot types?
- How well does H3 preserve identity, small text, logos and exact edits over many prompts and seeds?
- How much quality do FP8, INT8, NVFP4, pruning, Cache-DiT and reduced step counts lose?
- Does Context-IR improve average quality enough to justify the hosted dependency?
- What data was used for training, and what overlap or contamination checks were performed?
- What are the safety refusal, false-positive and false-negative rates?
The arena results are a meaningful first signal. They do not replace controlled multi-seed evals, worst-case analysis or a local-versus-hosted comparison.
Performance benchmarks: how fast can H3 run?
SGLang’s published H3 measurements are the strongest deployment data available. They report runtime and memory, not visual quality. Kingy did not reproduce them.
The most controlled table uses one eight-GPU NVIDIA B300 host. Each cell generated a 1344×768, 124-frame, 5.167-second clip at 24 fps with 50 steps, one request in flight and one measured request after warmup.
| Checkpoint | Precision | Encoder placement | Inference latency | Peak memory per GPU |
|---|---|---|---|---|
| FL2VA | BF16 | Auto/fold | 19.04 s | 83,578 MB |
| FL2VA | Online FP8 | Auto/fold | 18.03–18.04 s | 51,926 MB |
| Ref2VA | BF16 | Auto/fold | 29.12–29.13 s | 83,968 MB |
| Ref2VA | Online FP8 | Auto/fold | 27.12 s | 52,816 MB |
FP8 saved roughly 31–32 GB per GPU in that setup but improved latency by only about one second. SGLang labels online FP8 approximate and retains selected projections and heads at FP32. Memory, not speed, is the clear win.
Other published configurations are useful but should not be collapsed into one leaderboard because workloads, warmup coverage and topology differ:
| Hardware and topology | Published result | Practical reading |
|---|---|---|
| 4× H200, Ulysses4, BF16 | 74.38 s end to end; 94,290 MB/GPU on a warmed 1344×768 five-second T2VA run | Faster than TP2 + Ulysses2, but uses about 30 GB more per GPU |
| 4× H100 80 GB, TP2 + Ulysses2 | 13.25 s pipeline latency; 66.04 GB/GPU in SGLang’s topology comparison | Fastest of three measured H100 placements; preserve the exact recipe before comparing |
| 2× RTX 5090 32 GB, TP2 with layerwise offload | 559.67 s for a 50-step 1344×768 five-second request; 26.3 GiB/GPU | Lossless weights fit, but the host needs roughly 384 GiB RAM and the transfer cost is large |
| 1× AMD MI355X | 288.8 s denoise + 24.0 s decode; about 137.7 GB peak on a published 209-frame run | Single-GPU is possible on a 192 GB-class accelerator, not an ordinary Radeon |
| 1× AMD MI300X | 978.1 s denoise + 36.0 s decode; about 137.6 GB peak | Capacity exists; multi-GPU scaling is far more practical |
The B300 sweep measured one request per cell after one warmup, not a distribution across dozens of prompts. The H100 figure is also surprisingly far ahead of the published H200 run, a warning against ranking GPU models from rows with different measurement context. Treat each table as a recipe validation, not a universal throughput promise.
SGLang also reports Cache-DiT preview profiles on four H200s: mean inference fell from 75.10 seconds in lossless mode to 53.70, 30.23 and 25.81 seconds at progressively more aggressive settings. Same-seed video SSIM declined from 1.0 to 0.931, 0.818 and 0.794. Those metrics measure deviation from the lossless output, not whether the alternative looks good, and they do not score audio.
What hardware do you need to run MiniMax H3 locally?
Hardware requirements depend on which “H3” you mean, which checkpoint partition you need, whether quantization is acceptable, how much host RAM you have and how long you can wait.
Storage first
One original FL2VA or Ref2VA task partition is about 144.0 GB decimal, or 134.1 GiB. Both original partitions consume roughly 196 GiB after shared-file deduplication. A careless full repository checkout can approach 330 GiB because the repository carries original and Diffusers layouts side by side.
Budget at least 160 GB of fast NVMe for one original partition, 220 GB for both original tasks and 350–400 GB if you intend to mirror the full repository. NVMe matters when layers stream between storage, host RAM and the GPU.
Capacity arithmetic by precision
The following is planning arithmetic from the safetensors parameter counts, not a measured VRAM requirement:
| Hypothetical precision across all components | Pure weight floor | Why the real requirement is higher |
|---|---|---|
| BF16/FP16 | ~129 GiB | Official VAEs are FP32; activations, attention buffers and allocator workspace add memory |
| FP8/INT8 | ~64.5 GiB | Quantization scales, higher-precision modules and staging buffers remain |
| INT4 | ~32.2 GiB | No official all-INT4 quality-equivalent H3 pipeline exists |
The official files occupy 134.1 GiB because the VAEs are stored in FP32. The ComfyUI-recommended consumer stack is not “H3 at INT4.” It combines a pruned INT8 transformer, an NVFP4/AWQ text encoder, an FP16 video VAE and an FP32 audio VAE for about 39.55 GiB on disk.
Datacenter: the cleanest full-precision route
For a production service or serious batch use, start with the framework-published topologies rather than adding VRAM numbers on paper.
- Four H100 80 GB GPUs run SGLang’s lossless TP2 + Ulysses2 recipe at 66.04 GB peak per GPU. FSDP lowers the published peak to 57.01 GB; TP4 lowers it to 49.80 GB with a small latency cost.
- Four H200s can keep the full pipeline resident with Ulysses4. TP2 + Ulysses2 cuts per-GPU memory from about 94.3 GB to 63.5 GB at modest cost in the published run.
- Eight B300s have the most complete published sweep. Online FP8 cuts peak memory to roughly 52 GB per GPU.
- SGLang validates one to eight MI300X or MI355X GPUs. A single card needs about 138 GB of memory for the measured workload, so this means Instinct-class hardware.
- Diffusers documents a two-card split: two 80 GB GPUs for BF16 without streaming, or two 48 GB GPUs with INT8 loading.
Fast interconnect matters. Ulysses sequence parallelism divides the long token sequence but does not automatically shard every weight. Tensor parallelism and FSDP reduce per-card weight residency at the cost of collectives. High-bandwidth single-node NVLink/NVSwitch or tightly coupled Instinct systems are preferable to assembling cards across ordinary network links.
Consumer NVIDIA: possible, with compromises
The lowest specifically measured lossless consumer configuration is SGLang’s two RTX 5090s plus a 384 GiB-class host. It takes roughly 9.3 minutes for the published five-second, 50-step, 1344×768 request. That configuration preserves BF16/FP32 math but streams layers aggressively.
The Diffusers H3 branch documentation describes less expensive capacity routes:
- One 80 GB GPU with automatic CPU offload.
- One 24–32 GB consumer GPU with INT8 weight-only loading and block streaming. Expect around 75 GB of host RAM for the weights, plus operating headroom.
- A 12–16 GB GPU with the video VAE offloaded too and a smaller canvas such as 960×544.
These are capacity recipes, not speed recommendations, and the Diffusers integration was still on a feature branch rather than a normal release when checked. A 12 GB card may eventually complete a reduced-canvas job; that does not make it a pleasant H3 workstation.
ComfyUI’s official H3 tutorial is the more approachable consumer workflow at launch. Its model repository supplies pruned and quantized conversions. You still need roughly 42.5 GB of files for the recommended hybrid set, substantial system RAM for offload and patience. ComfyUI had not published a controlled H3 quality or speed comparison for those conversions at the time of review.
Consumer AMD, Apple Silicon and CPU-only
SGLang has verified H3 on AMD MI300X and MI355X through ROCm/AITER. That does not establish support for Radeon cards. Day-one community reports may help troubleshoot a specific Radeon, but they are not stable enough to turn into a purchase recommendation.
No official MPS, Metal or MLX recipe and no measured Apple Silicon result was available. A high-memory Mac may be able to hold quantized files in unified memory, but the required kernels, activation headroom and performance are unverified. H3 is not a supported Mac-local recommendation yet.
Pure CPU inference is a capacity thought experiment rather than a practical workflow. Even if a machine holds 150 GB of original weights or roughly 80 GB for an INT8/offload setup, it must evaluate a dense 33B diffusion transformer repeatedly over tens of thousands of tokens. Use CPU and NVMe as offload tiers, not as the main compute engine.
Hosted API pricing: often cheaper than buying the box
MiniMax’s pay-as-you-go pricing was straightforward when checked on August 3:
| Item | Price | Example |
|---|---|---|
| H3 768P output | $0.08 per output second | $0.40 for 5 s; $0.80 for 10 s; $1.20 for 15 s |
| H3 2K output | $0.13 per output second | $0.65 for 5 s; $1.30 for 10 s; $1.95 for 15 s |
| 768P-to-2K regeneration | $0.05 per output second | $0.50 to regenerate 10 s, with original inputs billed again |
| H3-Context-IR | $0.90/M input tokens; $3.60/M output tokens | Depends on the processed context |
| Image references | First five free; $0.04 each after | Nine images add $0.16 |
| Audio references | Free | Audio still must accompany image or video in Ref2VA |
| Video references | $0.08/s for 768P output or $0.13/s for 2K | A 10 s reference can cost as much as a 10 s output |
A regeneration task rebills the original input materials, so the cheapest apparent route is not always the final invoice. Video packages did not yet cover H3.
Hosted use also avoids the weight download, integration work and hardware bill while giving access to the two proprietary stages that make up MiniMax’s official 2K workflow. For many creators, local H3 makes sense only when privacy, customization, high volume or workflow control outweighs those costs.
The licence is a product limitation, not fine print
The MiniMax H3 Community License Agreement is restrictive enough to change deployment decisions. This is a plain-language reading, not legal advice.
- The licence applies worldwide except the European Union, United Kingdom, South Korea and United States. Use of the weights, derivatives or outputs in those excluded territories requires separate authorization from MiniMax.
- Commercial products and services generating more than $20 million in annual revenue need prior written authorization.
- A commercial product using H3 or its derivatives must display “MiniMax H3” prominently in its interface.
- H3, its derivatives and its outputs cannot be used to improve another AI model unless that model is H3 or an H3 derivative.
- A hosted product must bind users to comparable restrictions and maintain safeguards, reporting and enforcement mechanisms.
- Publicly disseminated machine-generated information or content must be clearly disclosed as machine generated under the acceptable-use policy.
- MiniMax says it claims no rights over outputs, while leaving responsibility for those outputs with the user.
The geographic exclusion is especially unusual. A Canadian studio may fit within the operative territory while an otherwise identical US studio does not. Distribution and output use across borders need legal review. “The weights are on Hugging Face” is not the same as permission to deploy them everywhere.
Limitations and failure modes to plan around
The open release is a base layer
Local H3 lacks the Context-IR component MiniMax calls critical to final quality and the Regenerate-2K stage used for the headline resolution. You can build your own prompt processor, and you can generate at other canvases through community frameworks, but neither is automatically equivalent to MiniMax’s hosted pipeline.
Semantic editing is not frame-faithful editing
Ref2VA treats source video as semantic context. SGLang warns that it may recompose the frame, reorder motion or cuts and alter timing. H3 exposes no denoising-strength control for preserving a source clip pixel by pixel. Use a conventional editor or a purpose-built transformation model when exact continuity matters.
Sparse attention is missing at launch
MiniMax trained H3 with sparse attention but did not ship that implementation in the first release. Full attention makes long, high-resolution and reference-heavy sequences expensive. This is one reason community capacity claims should not be read as throughput claims.
Separate task partitions increase storage and operations
FL2VA and Ref2VA use different transformer weights. A server that needs endpoint-frame generation and semantic multimodal reference generation must store and schedule both partitions or run separate services.
Fifteen seconds is a shot, not a finished video
H3 supports multiple shots inside one clip, but the output ceiling remains 15 seconds. Long-form production still needs shot planning, continuity control, selection, editing, audio mixing and often separate colour and text work.
Quality claims remain under-specified
MiniMax acknowledges that visual detail and multimodal understanding still have room to improve. There is no published failure rate for small text, branding, identity consistency, lip sync or non-English dialogue. Generate multiple candidates and review every frame and audio track before commercial use.
Hosted moderation can block legitimate work
MiniMax states that hosted text and media are automatically moderated and that false positives and false negatives remain possible. Teams with sensitive but lawful material need to test the hosted path before committing a production schedule.
Who should use MiniMax H3?
Use the hosted API if you are evaluating quality or shipping normal volumes
The API gives you the full pipeline, current 2K path and predictable per-second cost without a six-figure GPU host. It is the right baseline even if your long-term plan is local, because it lets you compare local H3-Base against the product represented in MiniMax’s demos and the Artificial Analysis arena.
Use local H3-Base if privacy or customization justifies the engineering
The weights make sense for research, fine-tuning, private source material, pipeline integration and high-volume workloads in permitted territories. Full-precision deployments belong on datacenter accelerators. Consumer operation is a quantization-and-offload project.
Wait if your workflow depends on portable open source
The territorial exclusions, hosted 2K dependency and missing sparse-attention code make H3 a poor foundation for a product that must deploy uniformly across the US, Europe and Asia. A more permissively licensed model with a completely local pipeline may be strategically cheaper even if its first-day arena score is lower.
MiniMax H3 FAQ
Is MiniMax H3 open source?
No. H3-Base is open-weight under MiniMax’s custom Community License Agreement. The full hosted system is not released: H3-Context-IR and H3-Regenerate-2K remain proprietary services. The licence is not OSI open source and excludes the US, EU, UK and South Korea.
Can MiniMax H3 run completely locally?
H3-Base can run locally and generates 768p-class video with stereo audio. The official full-2K workflow cannot run completely locally because Context-IR and Regenerate-2K are API-only at launch.
How much VRAM does MiniMax H3 need?
There is no single number. One original task partition contains about 134 GiB of weights. Published lossless recipes include four 80 GB H100s, four H200s and two 32 GB RTX 5090s with roughly 384 GiB host RAM. Diffusers documents INT8 streaming on 24–32 GB cards with about 75 GB host RAM, and reduced-canvas operation on 12–16 GB cards with more offload. Those smaller routes trade speed, precision or both.
Will MiniMax H3 run on a Mac?
No official or measured Apple Silicon route was available when this article was written. Unified memory may solve only the first problem, capacity. Backend support, optimized attention, VAE kernels and useful speed remain unverified.
Does MiniMax H3 generate audio?
Yes. H3 jointly generates video and native 32 kHz stereo audio, including dialogue, effects, ambience and music. Its audio-enabled Artificial Analysis results were among the strongest launch-day signals, but no public task-specific audio-quality or synchronization benchmark was available.
Is MiniMax H3 really a 2K local model?
The official local base model produces a 768-pixel short edge. MiniMax’s “2K” output uses a hosted regeneration stage that reprocesses the base result with the original context. The documentation does not define the marketing label as one exact pixel dimension.
How does H3 compare with Gemini Omni Flash and Seedance 2.0?
In the August 3 Artificial Analysis snapshot, H3 was in the top three across all six audio and non-audio pools. It ranked first for audio-inclusive editing, while its confidence interval overlapped Gemini’s. These are hosted-system preference results, not controlled local-model scores, and close ranks should not be described as decisive wins.
Can I use H3 commercially?
Only if the licence fits the territory and product. The standard licence excludes the US, EU, UK and South Korea, requires separate authorization above $20 million in annual product or service revenue, and imposes branding, downstream-use and safety obligations. Get legal advice before deployment.
Final verdict
MiniMax H3 deserves the attention. A model that can take text, frames, motion footage and audio references, then produce a coherent short video and stereo soundtrack in one pass, is materially different from a silent clip generator. Its first independent arena results place the hosted product among the strongest video systems available.
Four distinctions govern the verdict: the official 2K product is not fully local; the public weights are far larger than a lightweight 33B package; the independent scores evaluate the hosted pipeline; and the licence is not globally permissive. Each can change a buying or deployment decision without diminishing H3’s technical promise.
Evaluate the hosted API with your own multimodal shots first. Move to local H3-Base only when privacy, tuning or sustained volume is worth the hardware and operational cost, and only after confirming that the licence covers every place the model and its outputs will be used.
Official and primary sources
- MiniMax H3 launch article
- Official MiniMax H3 model card and files
- MiniMax H3 Community License Agreement
- MiniMax H3 API specification
- MiniMax pay-as-you-go pricing
- SGLang H3 deployment guide and performance measurements
- Diffusers H3 integration and memory recipes
- ComfyUI H3 tutorial
- Artificial Analysis video methodology
Kingy Launch Brief
Put the week’s verified AI launches in your inbox.
Get a source-checked briefing on consequential AI launches, with a clear try, watch or skip verdict. Beehiiv will ask you to confirm your address, then you can choose the subjects you want to follow.
Free · Choose your subjects · Double opt-in · Unsubscribe anytime