model release

MiniMax H3 Becomes First Open Video Model to Top an AI Video Ranking

TL;DR

MiniMax has released open weights for H3, a 33-billion-parameter video model that ranks first in Video Editing and second in Text-to-Video on Artificial Analysis — the first time an open model has topped a video generation category. The model accepts text, images, video, and audio in a single prompt, though its highest-resolution module remains closed.

3 min read
0

MiniMax has released open weights for H3, a 33-billion-parameter video generation model, making it the first open model to top a category on Artificial Analysis's video generation ranking. The model places first in Video Editing, second in Text-to-Video, and third in Image-to-Video, according to the benchmark site.

H3 is multimodal by design. It processes text, images, video, and audio together in a single request, generating clips between four and 15 seconds long with stereo sound. According to MiniMax's model card, a single prompt can combine up to nine reference images, three video clips, and three audio clips — a level of multi-input flexibility not typically available in open video models.

What's actually open — and what isn't

MiniMax did not release the full H3 stack. Two components remain closed: the 2K resolution module and H3-Context-IR, an internal system that translates prompts and reference material into a structured intermediate format before generation. Without H3-Context-IR, users running the model locally have to prepare that context manually, following prompting guides MiniMax has published separately.

The resolution gap is also notable in practice. Running H3 locally through ComfyUI currently tops out at 768p, well below the 2K output MiniMax's closed pipeline can produce. Users who want the model's full visual quality still depend on MiniMax's hosted service rather than the open weights.

What the open release does enable is fine-tuning. Developers can adapt H3 to custom footage, specific characters, or a particular visual style — something closed-model APIs generally don't allow. The license does carry a revenue cap, though: commercial use is only permitted for companies with less than $20 million in annual revenue, putting the model out of reach for larger commercial deployments without a separate agreement with MiniMax.

Pricing for MiniMax's hosted H3 service, including the closed 2K module, was not disclosed in the model card or accompanying materials.

Competing launch from ByteDance

The H3 release landed the same day ByteDance shipped Seedance 2.5, a closed video model that generates clips up to 30 seconds long with built-in audio — roughly double H3's maximum clip length. Seedance 2.5 is not available as open weights, and pricing details were not included in the source material.

What this means

H3's ranking result is the first concrete evidence that open-weight video models can compete with closed frontier systems on quality metrics, not just cost. But the release is only partially open: the highest-resolution module and the context-formatting system that likely contribute to H3's ranking performance stay behind MiniMax's hosted API. That split lets MiniMax claim an open-source milestone while preserving a commercial moat around the parts of the pipeline that matter most for output quality. For teams evaluating the model, the practical ceiling running it yourself is 768p — a meaningful gap from the 2K results driving its ranking placement. The $20 million revenue cap on commercial use further narrows who can actually build products on H3 without negotiating separately with MiniMax. Whether this becomes a template — rank-topping performance paired with partial openness — or a one-off will depend on how competitors like ByteDance and other Chinese labs respond in the coming months.

Related Articles

model release

Apple releases FastFrames-L, a 14B two-step video model distilled from Wan2.1-14B for interactive 480p generation

Apple has released FastFrames-L (14B) and FastFrames-S (1.3B), autoregressive two-step video diffusion models that generate 480p video at interactive latency. FastFrames-L is distilled from Wan2.1-14B, and the repository lists 35.4 ms total latency for FastFrames-S on one GPU.

model release

Tencent releases Hy4 Flash Base: 220B-parameter MoE, 12B active, 1M-token context

Tencent has published Hy4 Flash Base on Hugging Face: a 220B-parameter mixture-of-experts model with 12B active parameters per token, text and image input, and a 1M-token context. Tencent calls it its most cost-effective Hy model to date.

model release

HeyGen launches Avatar 6: native 2K video of a person from a single photograph

HeyGen introduced Avatar 6, an audio-driven avatar model that generates video of a specific person from a single photograph. HeyGen reports native 2K output and no fixed duration limit, with availability in its web app, mobile app and API from mid-November 2026.

model release

SenseNova-Loop-9B released: 9B parameters, 56 layer executions, 512K context

SenseNova has released SenseNova-Loop-9B, a roughly 9B-parameter multimodal model that runs 56 decoder-layer executions per forward pass from 32 distinct layers. It supports up to 512K tokens of context and, per the model card, needs about 15 GiB of GPU memory at 256K context with 4-bit weights and an 8-bit KV cache.

Comments

Loading...