MiniMax H3 generates five-second clips faster than playback on eight NVIDIA B300s

Why it matters

MiniMax's open-weight strategy is attracting outside infrastructure work that can lower serving costs and widen deployment options. Sol-H3 also shows how much a realtime claim depends on adapters, warm hardware and licensing.

Yan Junjie, founder, chairman, CEO and CTO of MiniMax, opened H3's weights and produced a concrete payoff: NVIDIA built an inference stack that generated a five-second, audio-synced clip in 1.653 seconds on eight B300 GPUs. MiniMax promoted the work in an official post on September 7th, three days after NVIDIA published the acceleration package.

Yan spent more than six years at SenseTime and became a vice president and vice-head of its research institute. He earned a doctorate in artificial intelligence from the Chinese Academy of Sciences' Institute of Automation and later conducted postdoctoral research at Tsinghua University. That research background is relevant here because MiniMax's open-model strategy depends on outside engineers being able to inspect H3, change its serving stack and publish improvements that MiniMax did not have to build alone.

The optimization work came from NVIDIA Research's Efficient AI Team and Singapore Lab. Core contributors include Yitong Li, Haopeng Li and Enze Xie. Their Sol-H3 release combines sparse attention, fused GPU kernels, faster communication between GPUs, parallel video decoding and cached conditioning inside one runtime.

NVIDIA released the Sol-H3 code under Apache 2.0. Reactor also published a FastH3 API for queuing, generating and playing clips. That gives developers a route to test the stack before configuring a multi-GPU machine.

Eight B300s and four denoising steps

NVIDIA's headline benchmark needs its full configuration attached.

NVIDIA's Sol-H3 benchmark generated 124 frames of 1344x768 video at 24 frames per second, with synchronized stereo audio, in 1.653 seconds. The result was the median of three requests after one warmup on eight NVIDIA B300 GPUs. Timing included text encoding, denoising and video-and-audio VAE decoding. It excluded model loading, compilation warmup and final MP4 encoding.

NVIDIA compared that result with a dense, 50-step H3 run that took 18.25 seconds. The resulting 11.04x figure does not come from runtime engineering alone. Sol-H3 uses a four-step FastH3 adapter, reducing the workload from 49 transformer forwards to four, and its fastest multi-GPU configuration uses approximate sparse attention alongside INT8 and FP8 data transport.

NVIDIA's own repository makes that distinction explicit. The default profile prioritizes speed and uses lossy acceleration, while a dense mode remains available when closer reference behavior is required. Performance on one B300 also falls below playback speed: NVIDIA reports 13.745 seconds for the same five-second output.

The benchmark demonstrated faster-than-playback MiniMax H3 generation on eight NVIDIA B300 Blackwell Ultra GPUs. By September 7, those were no longer NVIDIA's newest announced data-center platform: NVIDIA had introduced Vera Rubin and said deployments were underway. The benchmark does not establish that H3 runs faster than playback on a workstation, an RTX card or an uncached production request that must also load the model and encode the finished file.

NVIDIA has separately reported smaller inference gains for H3 on desk-side hardware. Its on-device Sol Engine work measured a 4.52x speedup on an RTX 5090 and 3.92x on a DGX Spark using inference-time kernel work, sparse attention and cross-step caching. Those five-second jobs still took 231.2 seconds and 181.3 seconds, respectively, because they retained the 50-step workload and ran different resolutions.

Outside engineers extend Yan's H3 bet

MiniMax released H3 on August 3rd, describing it as a 33-billion-parameter model that accepts text, images, video and audio and generates clips with native 32 kHz stereo sound. H3 supports outputs between four and 15 seconds at 24 frames per second, with a separate regeneration workflow for 2K output.

The five weeks since that release have turned H3 into a test of Yan's distribution strategy. MiniMax supplied the base model, while the acceleration stack layers in work from NVIDIA and a four-step adapter from Hao AI Lab at the University of California, San Diego. Reactor handles hosted access. Each participant gains something: MiniMax gets broader adoption, NVIDIA demonstrates demand for B300 systems, and Reactor gets a technically differentiated model for its API.

RuntimeWire previously reported that MiniMax was cataloging H3 integrations for local deployment, ComfyUI and multi-GPU serving. Vercel later cut H3 and H3 Max prices by 50% for a two-week promotion through its AI Gateway. Sol-H3 pushes the same strategy deeper into infrastructure. MiniMax is letting cloud platforms, research groups and serving frameworks compete to make H3 cheaper or easier to run.

That fits Yan's recurring emphasis on inference economics. RuntimeWire noted the same pattern when MiniMax priced an M3 business-email demonstration at $0.018, while cautioning that one visible run did not establish a repeatable cost. Sol-H3 offers a stronger technical artifact because NVIDIA published code, configurations and benchmark methodology. Its fastest number still belongs to a carefully selected hardware and model profile.

Open code, separately licensed weights

MiniMax's post described the acceleration as "fully open-sourced." The Sol-H3 code is Apache 2.0, but NVIDIA states that the model weights and third-party components keep their own terms.

H3's community license FAQ says open-weight use is temporarily restricted in the United States, European Union, United Kingdom and South Korea. Organizations in those regions can apply to MiniMax for authorization. MiniMax says its hosted API remains globally available because MiniMax can apply content and compliance controls to requests served on its own infrastructure.

The distinction shapes who can reproduce NVIDIA's work. Developers may inspect and modify the Apache-licensed acceleration code, while access to the underlying H3 weights remains governed by MiniMax's territory rules. Reactor's hosted service offers another path because users do not need to download the model before testing it.

For Yan, Sol-H3 demonstrates the upside of releasing a capable model into an active inference community. A research team at NVIDIA, an academic post-training group and an API provider assembled a faster deployment path in weeks. The next test is whether those gains survive ordinary production constraints, where hardware costs, cold starts, encoding and licensing carry as much weight as the fastest warm benchmark.