共有:

Open-Weight Video Model

Video generation just got
its first open-weight leader

Six months ago, video generation was still a closed-model game dominated by Veo, Runway, and Kling. China's MiniMax has rewritten that order with an open-weight release called "MiniMax H3." Running the whole pipeline on your own GPUs is now a real option — though for pure convenience, the closed players still have the edge.

AI Navigate Editorial2026.08.047 min read

VIDEO ARENA ELO 1241 1223 1204 1268 Veo 3.1 Runway Gen-4.5 Kling 2.5 MiniMax H3 First open-weight leader
01
What Happened

An open-weight model took
the top spot on Video Arena

A ranking closed players had owned just got an open-weight challenger.

On August 2, MiniMax announced on its official blog that its video model "MiniMax H3" had taken the overall top spot on the Video Arena leaderboard run by independent benchmarking firm Artificial Analysis — the first time an open-weight model had done so. Video Arena aggregates hundreds of thousands of anonymous head-to-head human preference votes, and for more than six months the top of that ranking had been occupied exclusively by closed players: Google's Veo, Runway's models, and Kuaishou's Kling.

MiniMax H3 posted an Elo score of 1,268, 27 points ahead of second-place Veo 3.1 (1,241). On the academic VBench-2.0 benchmark it scored 84.7% overall, and MiniMax says it edged out closed rivals specifically on temporal consistency and physical plausibility. The weights are released on Hugging Face under the MiniMax Model License (free commercial use for companies with under 100 million monthly active users), and inference is feasible on-premises with an 8×H100-class GPU cluster.

02

The handover, by the numbers

1,268
Video Arena Elo (new record)
84.7%
VBench-2.0 overall score
45B
Parameters (MoE-based DiT)
8×H100
Recommended self-hosted setup

Teams that can't send footage off-site
finally have a real choice.


03
Why Now

Why this matters now

What's different from six months ago — this is the first real crack in the monopoly.

Around Feb 2026 Veo, Runway, and Kling hold every top rank August 2, 2026 MiniMax H3 takes Video Arena's first open-weight top spot
FIG. From six months of closed-model dominance to the first open-weight handover

For the past six months, comparing video generation models has really meant comparing closed APIs against each other. Open-weight video models existed, but they trailed Veo and Runway by generations in quality and never came close to the top of the leaderboard. What this handover signals isn't just that one vendor caught up — it's that "running it on your own GPUs" has become a viable option without sacrificing quality. The assumption that cloud APIs were the only serious choice is starting to crack.

04
Who It Affects

Who it affects, and how

The payoff varies a lot depending on your role.

Designers

For unreleased product shots or client-supplied assets that can't leave the building, you can now generate rough-cut video in bulk on your own infrastructure and iterate freely without worrying about where the footage goes.

Marketers

Brands with sensitive creative can now prototype video variants without sending anything off-site. That said, for pure volume production where ease-of-use matters most, sticking with Veo or another closed model is still often the simpler call.

Engineers

Because the weights are public, you can fine-tune with LoRA on internal data and slot the model into existing inference stacks (vLLM-style tooling). Budget for the 8×H100-class cluster it takes to run it as a real infrastructure cost.

PMs & Leadership

You gain a vendor-lock-in-free option for bringing video generation in-house, which is real leverage in cost negotiations. But GPU procurement and operations add fixed costs, so the break-even point versus API billing depends heavily on your volume.

05
Closed vs Open

What actually changes versus the closed players

Closed players (Veo 3.1 / Runway Gen-4.5 / Kling 2.5)MiniMax H3 (open-weight)
API access only, weights not releasedWeights public, runs on your own GPUs
Generated data assumes a cloud round-tripCan run fully in-house or on-prem
Video Arena Elo 1,204–1,241Video Arena Elo 1,268 (leader)
No modifying or fine-tuning the model itselfCustom tuning possible via LoRA and similar
Polished UI, low setup effortRequires GPU procurement and ops know-how

06
What's Next

What happens next, and recommended actions

MiniMax H3's top spot is a snapshot as of August 2, and Google, Runway, and Kuaishou could all plausibly ship counter-releases within weeks that retake the lead. Model turnover on video leaderboards has been fast this past year, so the ranking itself matters less than the structural fact it proves: an open-weight model can compete at the top. That's the part likely to matter longer than any single leaderboard position. (1) If you have workloads that can't send footage off-site, it's worth standing up a small-scale GPU proof-of-concept with MiniMax H3 now. (2) If ease-of-use and volume are what you care about, there's no need to change how you use Veo or other closed models yet. (3) Given how quickly the ranking could shift again in the coming weeks, avoid over-investing your pipeline in any single model for now.

07
Risks & Limits

Counter-view, risks, and limits

Don't overread this. First, Video Arena is a human-preference leaderboard, and scores shift with the composition of voters and the mix of prompts used. Second, self-hosting requires real upfront investment — an 8×H100-class cluster and staff who can maintain an inference stack — so "open" doesn't mean "instantly usable by anyone." Third, the MiniMax Model License requires a separate commercial agreement for large operators (100 million-plus monthly active users), so this isn't unconditionally open source. On convenience and surrounding tooling, Veo and Runway still have the edge.