MiniMax releases H3 open weights with day-zero vLLM-Omni support

• MiniMax H3 weights are live, with the H3 availability post and a Hugging Face pointer both landing on the public model release.
• H3 folds text-to-video, image-to-video, first/last-frame control, reference-to-video, and in-place editing into one multimodal model, according to ComfyUI's support post.
• Day-zero infra is unusually complete: vLLM-Omni's post covers serving, ComfyUI's post covers graph workflows, SGLang Diffusion's post covers local serving, and fal's post covers hosted endpoints.
• The open release has a hosted middle: the Hugging Face model card says H3-Context-IR is not included, while a model-card link shared in a Hugging Face pointer confirms where the release landed.
• Access is region-sensitive: one license concern read the license as blocking USA/EU/UK/Korea use, and teortaxesTex's screenshot captured a MiniMax reply tying US limits to Hollywood litigation.

The model card hides the best engineering caveat: local H3-Base is the 768p path, while Context-IR and 2K regeneration still lean on hosted APIs. The vLLM recipe is Christmas come early for serving nerds, with /v1/videos, partition limits, B300 timings, and the 8x64 GB OOM warning all in one place. The ComfyUI post says its path shrinks the footprint from 123.6 GB to 42.5 GB; one early license read found the catch before most people had finished downloading.

MiniMax's launch post framed H3 as a general-purpose multimodal generation model that reads text, images, video, and audio, then generates video with native stereo sound up to 15 seconds at 2K. By August 3, the weights were on Hugging Face, with the release split into task-specific checkpoints.

The model card gives the practical shape:

• Output duration: 4 to 15 seconds.
• Frame rate: 24 FPS.
• Audio: 32 kHz stereo.
• Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, plus other supported dimensions.
• H3-Base-FL2VA: zero images falls back to text-to-video; one image handles first-frame or last-frame generation; two images handle first-and-last-frame generation.
• H3-Base-Ref2VA: up to 9 images, 3 video clips, and 3 audio clips, with a 12-file mixed-input cap.
• Reference audio rule: audio cannot be the only reference input in Omni Reference mode.

ComfyUI's launch note put the workflow names in creator terms: prompt-only T2V, image-to-video, first-and-last-frame, reference-to-video, and in-place editing ComfyUI's support post.

The released H3-Base architecture packs encoded modalities into one multimodal sequence before the H3-Omni-Transformer predicts video and audio latents. The Hugging Face model card lists the parts that matter for implementers:

• Text and vision encoder: H3-Encoder uses Qwen3-VL-32B and passes hidden states from layer 50 into the Omni-Transformer.
• Visual latents: H3-VisualVAE uses f16t4d24 compression, then patchifies visual latents with a 1 x 2 x 2 patch size.
• Audio latents: H3-AudioVAE compresses 32 kHz stereo audio into latent tokens at 40 Hz per channel.
• Transformer: H3-Omni-Transformer is a 33B dense, single-stream Transformer.
• AdaLN branches: roughly 13B parameters sit in AdaLN-related branches, and the model card says their modulation outputs can be precomputed and cached for inference-only deployment.
• Sparse attention: H3 supports sparse-attention training and inference, but the initial open release provides full-attention inference only.

That Qwen3-VL dependency is why the encoder is not a footnote. In the vLLM recipe, the Qwen3-VL layer-50 text/vision encoder alone is listed at 51.5 GB BF16.

MiniMax's full H3 system has three named modules: H3-Context-IR, H3-Base, and H3-Regenerate-2K. The model card says H3-Context-IR is a hosted preprocessing and orchestration system that parses multimodal instructions, cross-modal associations, temporal relationships, and underspecified details into a Context Intermediate Representation.

The local release does not include that module. MiniMax says H3-Context-IR depends on multiple hosted models and services, and the release provides APIs, tutorials, and prompting guidance for reproducing the official workflow.

2K is also a multi-stage path. H3-Base generates 768p audio-video, then H3-Regenerate-2K feeds that result plus the original context back into H3 to produce 2K output; the model card says this regeneration module is not yet open-sourced.

MiniMax's launch post adds one useful number: its Contextual Omni Representation pipeline can require around 100K tokens of inference on source material, distilled down to roughly 4K tokens on average.

The vLLM team shipped day-zero H3 support through vLLM-Omni. The vLLM recipe serves H3 through an OpenAI-compatible /v1/videos API, with /v1/videos/sync returning the MP4 directly and /v1/videos using async job polling.

The serving shape is explicit:

• Install: vLLM 0.26.0+, vLLM-Omni nightly, Docker or pip.
• Checkpoint location: serve from a local path after downloading the gated Hugging Face checkpoint.
• Partitions: FL2VA handles t2va and fl2va; Ref2VA handles ref2va.
• Server model: one server process loads one partition, so switching between FL2VA and Ref2VA means restarting against the other path.
• Components: 66.3 GB joint video/audio DiT, 51.5 GB Qwen3-VL encoder, roughly 10 GB video VAE, roughly 0.6 GB audio VAE.
• Validated four-GPU path: the recipe reports 86.96 seconds client end-to-end for an 8.7 second FL2VA workload on 4x B300.
• Memory cliff: the recipe says 8x64 GB will OOM under the analyzed Ulysses-only setup because the 66.3 GB DiT is replicated per rank.

The MP4 output is H.264 video with synchronized native stereo audio, generated jointly rather than muxed from a separate audio model, according to the same vLLM recipe.

ComfyUI shipped native H3 support on day zero, with templates for I2V, R2V, and T2V workflows. Its launch post says H3 requires ComfyUI 0.30.0 or newer.

The ComfyUI-specific engineering is the memory story:

• Modulation pruning: ComfyUI says H3's modulation weights are about 40% of total parameters and can be pruned into a functionally equivalent lookup table.
• Quantization: the weights ship with int8 convrot quantization.
• Kernels: custom kernels reduce peak VRAM during inference.
• Footprint: ComfyUI reports a 66% reduction, from 123.6 GB full precision to 42.5 GB for the smallest variants.
• Consumer GPU claim: dynamic VRAM offloading lets the optimized path run locally on an RTX 3060, according to ComfyUI's post.

That is the biggest practical split between the day-zero stacks: vLLM documents the high-throughput server path, while ComfyUI documents the smallest local graph path.

fal put three H3 endpoints live at launch: text-to-video, image-to-video, and reference-to-video. Its endpoint pages linked from fal's post price 2K generation at $0.26 per second for text-to-video and image-to-video, with reference assets adding their own per-second or per-image costs.

SGLang Diffusion shipped local serving support too. Its launch note claimed H3 can run locally on 2x RTX 5090 or 1 RTX 6000, and named NVIDIA Blackwell/Hopper plus AMD MI355X and MI300X as supported hardware.

Other launch surfaces showed up fast:

• vLLM-Omni: OpenAI-compatible video endpoint, sync or async vLLM-Omni support post.
• ComfyUI: native nodes and graph workflows ComfyUI support post.
• fal: hosted T2V, I2V, and R2V endpoints fal endpoint post.
• SGLang Diffusion: local deployment across NVIDIA and AMD hardware SGLang Diffusion support post.
• Vercel AI Gateway: the thread around AI Gateway's H3 code sample used the model slug minimax/minimax-h3 inside generateVideo().

fal's demos also showed why native audio matters. Its soundtrack demo described a prompt that placed a low beat at 3 seconds, jazz bass at 6 seconds, a sax riff at 10 seconds, and a freeze on a tense chord, then generated a 15-second title sequence from five still images.

The release landed with a licensing wrinkle. One community read said the license forbids usage in the USA, EU, UK, and Korea, and said users in the USA may not even download the model license concern.

The screenshot in teortaxesTex's post includes a reply attributed to RyanLee at MiniMax saying the US carve-out stems from ongoing generative-video copyright litigation with major Hollywood studios. The same reply says US-based persons can submit formal licensing requests, with licenses issued to applicants that implement compliance controls and guardrails aligned with US legal, compliance, and regulatory requirements.