← Back to feed

MiniMax H3 was added to Arena’s Video Edit leaderboard on August 13, landing as the latest state-of-the-art entry in a category ByteDance has dominated since Seedance 2.0 arrived. The model is competent: oxen.ai’s hands-on review called it “state of the art across 3 core video generation tasks” on the Arena leaderboard, and community benchmarkers are already building on it at speeds that don’t happen for incremental updates.

But reading MiniMax H3 as a clean open-weights story misses the more significant detail buried in the license.

What H3 Is

H3 is a 33B-parameter diffusion transformer with Qwen3-VL-32B as its visual encoder. It shipped July 31, with partial weights following on August 7. The model takes text, up to nine reference images, three reference videos, and three audio clips simultaneously: then outputs 4-15 second clips at 768p locally or 2K through the hosted API pipeline. Native 32 kHz stereo audio ships with every generation, which no other video model in the current crop matches.

The architecture splits into three stages: Context-IR (contextual processing), H3-Base (the generation model), and Regenerate-2K (the upscaler). Only H3-Base is in the public weight release. The other two stages remain hosted-only.

Two generation modes cover the main use cases. FL2VA (First/Last Frame to Video/Audio) anchors the clip to zero, one, or two image endpoints and interpolates. Ref2VA handles style and subject consistency across up to 12 mixed-media inputs. The community has already built a 4-step sampling LoRA that delivers roughly 5x speedup and a quantized 8GB VRAM version for users who can’t afford A6000-level hardware. Running on Apple Silicon works via MLX; the first public tutorial appeared the week of August 4.

API pricing: $0.13 per second for 2K resolution output, $0.08 per second for 768p, with image input at $0.04 per image and video input at $0.13 per second.

The Licensing Structure

This is where H3 breaks from every comparable open-weight video model released to date.

The MiniMax H3 Community License identifies the United States, the United Kingdom, the European Union, and the Republic of Korea as “Excluded Territories.” Users in those regions can download the weights, but are required to submit an application to MiniMax before using them under the same terms that apply elsewhere.

The rationale given is that MiniMax needs to “ensure usage is lawful, responsible, and without infringing any rights.” No approval timeline is specified in the license.

Additional restrictions apply universally: the license bars all users from distilling another model on H3’s outputs, requires commercial products to prominently display the H3 model name, and prohibits uses that may harm minors, interfere with elections, or violate local law.

The effect is a two-tier open-weight structure: users in China and most of the rest of the world get immediate commercial access; users in Western markets and South Korea get a gated process with undefined latency.

What This Sets as Precedent

MiniMax H3 is not the first Chinese AI model to implement territorial restrictions, but it is the most explicit example applied to an open-weight video generation model at this capability level. Kimi K3 shipped its 2.8-trillion-parameter weights to Hugging Face last month under permissive terms. DeepSeek V4 Pro shipped MIT. Qwen3.8-2.4T shipped Apache 2.0.

H3 takes a different posture: one that looks designed to satisfy Chinese regulatory requirements around technology export while still pursuing the distribution benefits of open weights outside China.

For enterprises evaluating H3 for production use in the US or EU, the application requirement adds procurement risk that doesn’t exist with any Anthropic, Google, or Meta video model. For researchers, it raises questions about reproducibility, fine-tuning rights, and whether community extensions remain permissible after MiniMax reviews an application.

DeepLearning.AI’s writeup noted that “despite the restrictive license and unusual territorial restrictions, MiniMax H3 is clearly a top video generation model.” That sentence captures the tension precisely. The model earns its Arena spot on capability grounds. Whether Western users can build on it commercially without friction is a separate question that the license currently leaves open.

Current State

H3-Base FL2VA and H3-Base Ref2VA weights are publicly downloadable. Context-IR and Regenerate-2K remain hosted-only, meaning local runs top out at 768p without a custom upscaling pipeline. MiniMax’s hosted API delivers the full 2K output at the rates above and is described as globally available with safety filters applied.

The model is on Arena Video Edit. Whether its benchmark position translates into production adoption in markets where the licensing process adds friction is the test that follows.