MiniMax H3 guide: specs, license and prices, updated weekly

Updated 9 Nov 2026 · next update 16 Nov 2026

Key facts as of 26 Sep 2026

• The model: MiniMax H3 is a general-purpose omni-modal system built on a 33B-parameter dense single-stream Transformer, with about 13B of those parameters in AdaLN branches that can be skipped at inference because their outputs are cached. (source)
• Output: up to 15 seconds at 2K, with the shorter side at 768 pixels by default and 32 kHz stereo audio generated natively; 2K comes from an H3-Regenerate-2K pass that feeds the 768p result back through the model rather than from a separate super-resolution module. (source)
• Release date: the community license carries the date the release is tied to, stating "MiniMax H3 release date/License date: August 2, 2026". (source)
• License: royalty-free worldwide except the European Union, the United Kingdom, the Republic of Korea and the United States; commercial products and services that earn more than $20 million a year in revenue need MiniMax's separate written authorisation, and a commercial product using H3 must display "MiniMax H3". (source)
• Territories: MiniMax's grant is limited to the applicable territory and it invites anyone in an excluded territory to contact it for a license. (source)
• What the open weights do not include: H3-Context-IR, the stage that parses an instruction across modalities, needs hosted models and services and is not in the open release; H3-Regenerate-2K, the module that produces 2K, is not yet open-sourced either, so local weights give 768p and MiniMax's API is the route to 2K. (source)
• Attention: the initial open release runs full attention only, with sparse attention promised in a later update. (source)
• Rate at fal: $0.05 a second at 480p, $0.06 at 768p, $0.13 at 2K and $0.16 at 4K. (source)
• Rate at Higgsfield: the API console listed MiniMax H3 at $0.091 a second on 26 September 2026, 30% off a $0.13 list. (source)
• H3 Max: Higgsfield added MiniMax H3 Max on 31 August 2026, a post-trained variant tuned for prompt adherence and aesthetics that caps at 768p where the base model reaches 2K. (source)

MiniMax H3 is now the base layer under several other makers' models - Pika's six-step H3 Flex, Pruna's P-Video-2-Pro and Utopai X are all post-trained on it - and it sits third on Artificial Analysis's human-preference text-to-video board at 768p, with its own weights still open under a license that excludes the EU, the UK, South Korea and the United States. (source)

What MiniMax H3 is

MiniMax H3 is MiniMax's general-purpose multimodal generation model, announced on 31 July 2026 with weights released under a community license dated 2 August 2026. It generates video with native stereo sound, up to 15 seconds, at up to 2K, and it takes text, images, video and audio in the same context, which is the point of the design: MiniMax says H3 abandons the split between text-to-video, image-to-video, motion reference, voice reference and editing in favor of one model that follows instructions expressed in natural language (launch post).

The architecture is published. The model card describes a 33B dense single-stream Transformer with roughly 13B parameters in AdaLN branches, which can be cached and left out of an inference-only deployment; text is encoded by an H3-Encoder that uses the full pretrained weights of Qwen3-VL-32B and passes hidden states from its 50th layer; and audio is compressed by an H3-AudioVAE at 40 latent tokens a second per channel. Video and audio latents are predicted jointly by the H3-Omni-Transformer and decoded into picture and stereo sound. The card's deployment example is SGLang on four GPUs, and it points to a separate deployment guide for other configurations.

Three parts of the system are not in the open weights. H3-Context-IR, the stage that parses an instruction and associates it across modalities, relies on hosted models and services, so it is excluded and MiniMax offers an API that reproduces its behavior. H3-Regenerate-2K, the module that turns a 768p result into 2K, is described in the card as not yet open-sourced, with a MiniMax API for validating official results. And the initial release runs full attention: sparse attention, which the card says reduces the cost of long multimodal sequences, comes in a later update.

MiniMax's own framing of the price is that H3's per-second cost is less than a third of mainstream models at 2K and less than half of their 720p price at 768p. Both are the vendor's claims about its own model rather than measured comparisons.

What MiniMax H3 costs

• Open weights: free to download under the MiniMax H3 Community License, and the local model reaches 768p. Higher resolutions require MiniMax's own API, and the license restricts where the weights may be used at all.
• fal: published rates are $0.05 a second at 480p, $0.06 at 768p, $0.13 at 2K and $0.16 at 4K, across durations of 5 to 15 seconds and seven aspect ratios.
• Higgsfield: the API console listed H3 at $0.091 a second on 26 September 2026, marked 30% off a $0.13 list.
• MiniMax's own app: Hailuo AI lists H3 in its tool set and quotes it by the second: 2K video from $0.081 a second and 768p video from $0.047 a second. (source)
• H3 Max resellers: Pixazo cut the price of H3 Max on its API to 4 cents a second on 23 September 2026, as SLOP TV reported.

There is no per-token arithmetic to do here, which makes H3 easier to budget than models billed on output area. The resolution you pick is most of the cost: 768p to 2K is more than double the rate at fal, and the open weights stop at 768p.

What MiniMax H3 does

Video with sound in one pass. Up to 15 seconds at 2K, with native stereo audio at 32 kHz. MiniMax says voice, sound effects and music are modeled together rather than as separate tasks, and that multi-shot structure is native rather than assembled afterward.

Instruction following across modalities. The prompt in MiniMax's own example is a relationship between inputs rather than a description: "Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3." The model is meant to work out what that means.

Reference and edit. The launch post lists generalised reference and editing across image-to-image, image-to-video, audio-to-audio and audio-video-to-audio-video, which is the same generality claim stated as a capability list rather than as a set of modes.

Two limits worth knowing. Fifteen seconds is the ceiling, against the 30 seconds ByteDance claims for a single pass of Seedance 2.5, so longer work is a stitching job. And 2K output is a second pass over the 768p result rather than a single higher-resolution generation, which is why the model reuses the original context instead of a super-resolution module: MiniMax says this recovers small text and fine detail that upscaling can only guess at. That second pass is one of the parts that is not in the open weights.

Where it runs. Open weights for local use, the MiniMax API, Hailuo AI, fal, Higgsfield, and inside other products' model menus, including Perplexity's Computer workspace, which SLOP TV reported adding H3 and Seedance 2.5 in September 2026.

Who MiniMax H3 is for

Three distinct buyers. Teams that need video generation on their own hardware are the ones the open weights are for, and the license decides whether they qualify: deployments inside the European Union, the United Kingdom, the Republic of Korea and the United States are outside the granted territory, and MiniMax invites anyone there to contact it for a license. A company whose commercial products and services earn more than $20 million a year has to get MiniMax's written authorisation before using H3 commercially, and any commercial product built on it must display "MiniMax H3" on the interface.

Studios and developers who want an endpoint rather than a deployment compare per-second rates, where the resolution ladder decides the cost. And creators who want a finished clip buy it inside an app, where the subscription allowance is the constraint rather than the rate.

The license is the thing to read before building on H3. It is royalty-free and allows commercial work and derivative models, but it limits use to the applicable territory, which excludes the European Union, the United Kingdom, the Republic of Korea and the United States, and it adds a revenue threshold and an attribution requirement on top.

The MiniMax H3 diary

Dated entries, newest first. Each entry is added once and never rewritten; corrections are made in a new entry.

9 Nov 2026

Replicate published a comparison of MiniMax's three H3 video models on 3 November, with prices it took from its own model pages on 31 October: H3 Highspeed at $0.06 per output second at 768P, H3 Lite at $0.04 at 768P and $0.025 at 480P, and H3 at $0.08 at 768P and $0.13 at 2K, which makes a 10-second 768P clip $0.60, $0.40 or $0.80 depending on the tier. Replicate's own advice is to iterate on Highspeed, batch on Lite and pay for H3 only for 2K or a hard shot, and it notes that Lite is the model whose output most needs checking for a swapped object or a dropped character. The rates are the reseller's, not MiniMax's. Source: replicate.com.

Video Rebirth released HyperFlow H3 Flash on 6 November, a compressed derivative of MiniMax H3 that generates video with synchronized stereo audio in four, six (the default) or eight sampling steps from 13.7B parameters against the base transformer's roughly 31B, by dropping three transformer blocks and factoring 102 linear layers and then quantizing 180 layers to INT8. The code is Apache 2.0 on GitHub and the maker reports a 5.2-second clip in 8.7 seconds on four B200s, 36 seconds on one H200 and 92 seconds on one RTX 5090, with the smallest GGUF build at 9.1 GB running on 24GB and 32GB consumer cards with 16GB of system memory; 16GB cards are not measured. Because it is a derivative of MiniMax H3, the H3 community license and its territorial limits flow through to the release. Source: github.com.

2 Nov 2026

On 29 October Pika launched H3 Flex, a family it post-trained on MiniMax H3 to generate video and synchronized audio in six denoising steps instead of the base model's 50, publishing weights on Hugging Face, code on GitHub, and support in its API, ComfyUI and Diffusers. Pika says it ranks first among open-source text-to-video models, supports 480p to 4K and offers five separately editable audio stems - dialogue, score, SFX, foley and ambience - and its own comparison table prices it from $0.01 an output-second, against $0.08 for MiniMax H3 and $0.103 for Seedance 2.5. Every ranking and price in that table is Pika's own measurement of its own model. Source: pika.art.

On Artificial Analysis's text-to-video board, as read on 2 November, MiniMax H3 at 768p sits third at 1161 Elo behind Wan 3.0 on 1175 and Utopai X on 1166, and Pruna's H3-based P-Video-2-Pro entered at 1125 Elo in eleventh place. The board fits its scores from pairwise human-preference votes and lists five models added in the last 30 days, P-Video-2-Pro, FLUX 3 Max, MiniMax H3 Max among them, so a placement is a snapshot rather than a fixed rank. Source: artificialanalysis.ai.

26 Oct 2026

On 20 October the FastVideo team, working with Nuva Lab, Reactor and Nvidia, open-sourced FastH3 Omni, a distilled MiniMax H3 for reference-guided video with sound that runs in eight transformer evaluations instead of the base model's 50; it reports 5.6 to 11.9 times lower end-to-end latency across 27 matched workloads on eight GB200 GPUs, reaches real-time generation for 768p clips of 5, 10 and 15 seconds on its own serving stack, ships NVFP4 checkpoints for a DGX Spark or RTX 5090, and offers ComfyUI support. Every latency figure is the team's own measurement of its own serving recipe, not an independent benchmark of output quality. Sources: haoailab.com, huggingface.co, sloptvnews.com.

Taobao's content AI team published PTD LoRAs for H3: per-step low-rank adapters trained with Parametric Trajectory Distillation that cut the reference-image-to-video-and-audio path to eight steps and to four, alongside a sibling release trained with PTD plus Reference-Asymmetric On-Policy Distillation that ships eight- and six-step variants and claims better background plausibility, sharpness and realism than both the plain PTD checkpoints and the base model at 40 evaluations. Neither release has an independent quality benchmark. Sources: huggingface.co, arxiv.org.

Pruna's P-Video-2-Pro-Reference, a model built on MiniMax H3, now lists $0.015 a second at 480p in its cost mode, with 768p native output and 1080p, 2K and 4K generated by upscaling the 768p result; its documentation ties use to the H3 community license, including that license's acceptable-use policy. Sources: docs.api.pruna.ai, pruna.ai.

19 Oct 2026

On 13 October MiniMax published its write-up of the Japan IP.AI Co-Creation Conference it co-hosted with KAGAMIAI in Shibuya on 11 September, where it unveiled MiniMax H3 IP Edition: H3 paired with officially licensed Japanese IP and delivered through KAGAMIAI's Kagami Gate, an infrastructure that attaches approved assets, generation rules, commercial terms and usage reporting to each property before creation and monitors compliance during generation. MiniMax says more than 60 companies have since expressed interest and that availability depends on each partner's terms; the same post reports the Captain Tsubasa x PixVerse pilot's numbers - 126,360 creators across more than 195 countries producing 329,314 pieces over 45 days, with 0.14% of output violating program policies. Sources: minimax.io, x.com.

H3's open weights kept collecting community tooling: FastH3-Omni, a variant that adds reference images and audio to the model, and DeepBeepMeep's WanGP checkpoint set, which logged its H3 additions on 15 October - a rank-128 accelerator profile for the FL2VA path and a cut-down 12-layer LightVAE Lite decoder in BF16 and INT8. Sources: huggingface.co, sloptvnews.com.

12 Oct 2026

MiniMax has opened M Plan as the successor subscription to Token Plan: three tiers at $22 (Go), $55 (Explore) and $132 (Build) a month, with the H3 video model included on Explore and Build but not Go, and the first month 50% off through 14 October 2026. Sources: platform.minimax.io, platform.minimax.io.

26 Sep 2026

The hub opens. This first entry records the news the page starts from.

On 26 September the Higgsfield API console listed MiniMax H3 at $0.091 a second, marked 30% off a $0.13 list. Source: Higgsfield console.

On 23 September Pixazo cut the price of MiniMax H3 Max on its API to 4 cents a second, SLOP TV reported the same day. Source: SLOP TV.

On 22 September SLOP TV reported HelloGen's unlimited generation on MiniMax H3 at $19.90 a month with a floor of two priority renders a day, and Perplexity's addition of H3 to its Computer workspace at $20 and $200 a month. Sources: SLOP TV, SLOP TV.

On 31 August Higgsfield added MiniMax H3 Max, a post-trained variant tuned for prompt adherence and aesthetics that caps at 768p where the base model reaches 2K. Source: Higgsfield changelog.

MiniMax announced H3 on 31 July 2026 with native stereo audio and up to 15 seconds at 2K, and released the weights under its community license dated 2 August 2026. Sources: MiniMax, license.

What changed this week

Against the 2 Nov 2026 edition: Replicate's three-tier per-second ladder for H3 and Video Rebirth's HyperFlow H3 Flash are the week's two structural additions, the first a reseller's price list from $0.025 to $0.13 a second and the second an open-weights compression of H3 to 13.7B parameters that runs on consumer cards. MiniMax's own newsroom, read on 9 November, still shows nothing newer than the 12 October Japanese IP post, and the fal and MCP rates in the key-facts block were not re-quoted this week. Our own coverage this week added Viggle's two open H3 remake LoRAs, H3 Dials steering in ComfyUI, and Microsoft's Foundry listing of H3 Max and H3 Max Turbo through fal.

SLOP TV's MiniMax H3 coverage

• H3 Dials turn two prompts into a MiniMax H3 slider (8 Nov 2026)
• Video Rebirth cuts MiniMax H3 to 14B in six steps (7 Nov 2026)
• Microsoft Foundry adds five AI video models through fal (7 Nov 2026)
• Viggle Remake reshoots video with two open MiniMax H3 LoRAs (5 Nov 2026)
• How to use ComfyUI 0.44.0's dubbing, lip-sync and camera nodes (4 Nov 2026)
• The 5 open-source AI video tools that shipped, ranked by impact (3 Nov 2026)
• Pruna's P-Video-2-Pro lands 11th on the AI video leaderboard (31 Oct 2026)
• AI video discounts end today as per-second prices rise (31 Oct 2026)
• Pika H3 Flex cuts video generation from 50 steps to six (30 Oct 2026)
• Pika H3 Flex ships open weights under a license that bars the US (30 Oct 2026)
• MiniMax H3's 2K finish is now an open 3-step pass (29 Oct 2026)
• Utopai Studios raises $60M at $1B, The Information reports (26 Oct 2026)
• MiniMax H3 cuts sampling to 4 steps with two open releases (25 Oct 2026)
• The 8 cheapest AI video models per second, from $0.015 (22 Oct 2026)
• Omni reference: what AI video models now take as input (22 Oct 2026)
• Pruna's P-Video-2-Pro-Reference starts at $0.015 a second (18 Oct 2026)
• FastH3-Omni adds reference images and audio to MiniMax H3 (15 Oct 2026)
• 6 open-weight AI video models that make sound, ranked by GPU (13 Oct 2026)
• Seedance 2.5 tops an AI video physics test at 57.76 out of 100 (12 Oct 2026)
• How to run MiniMax H3 on a single 12GB graphics card (9 Oct 2026)
• Protoface opens its API to 40+ models from $0.003 a second (9 Oct 2026)
• How to run MiniMax H3 on an 8 GB graphics card (8 Oct 2026)
• HeyGen's first video model sells from 1 cent a second (2 Oct 2026)
• Pixazo cuts MiniMax H3 Max API price to 4 cents a second (23 Sep 2026)
• fal.live comes back with AI-generated prompts viewers can upvote (23 Sep 2026)
• Perplexity Computer adds MiniMax H3 and Seedance 2.5 video (22 Sep 2026)
• HelloGen sells unlimited AI video on one model for $19.90 (22 Sep 2026)
• The 7 Cheapest AI Video APIs Right Now, Ranked by Price Per Second (20 Sep 2026)
• Google's Veo, Alibaba's Wan and Pruna AI Are Now Fighting Over Cents Per Second (20 Sep 2026)
• OpenAI's Sora API Shuts Down Sept. 24: Where Its 4 Biggest Jobs Go Next (20 Sep 2026)
• MiniMax Open-Sourced H3 to Buy Distribution. The Market Paid, Then Took It Back (19 Sep 2026)
• Pika Relaunches as a Multi-App Studio With 3-Minute Videos & Automatic Model Routing (19 Sep 2026)
• MiniMax H3 Add-Ons Are Piling Up on Hugging Face: 8-Step LoRAs, a WASD World Model & More (19 Sep 2026)
• MiniMax H3 License Explained: Who Can Run the Open Weights & What the 4 Excluded Regions Must Do (19 Sep 2026)
• 5 Best Open-Weights AI Video Models to Run Yourself, Ranked: LTX-2.5 Tops Wan 2.2 (19 Sep 2026)