MiniMax-H3 surpasses 4.8 million downloads and LTX-2.5 hits 1.5 million: open-weight video generation becomes a commodity no one can lock in

MiniMax-H3 surpasses 4.8 million downloads and LTX-2.5 1.5 million: open-weight video generation becomes a commodity no one can lock down

🔎 AI video just crossed a threshold — and this time, the numbers are public

There are shifts you can sense coming, and others you can measure. This one is measurable: according to the Hugging Face trending leaderboard and the monitoring radar that spotted it, MiniMax-H3 has become the most downloaded video model on the platform, with 4.8 million total downloads and 5,293 likes on the counter (September 2026).

On the other side, the closed APIs — Veo, Kling, Seedance — continue to sell quality at a per-second rate. But a second front is opening up: Lightricks' LTX-2.5 passed 1.5 million downloads about a month after its open-weights release, and the ComfyUI ecosystem integrated both models on their release day.

Video generation is thus replaying the open LLM scenario, scene by scene: the big labs lock things down through quality, the community catches up through open weights, quantizations, and local tools. What was once a researcher's niche becomes a commodity — that is, a building block no one can lock down anymore.

Key Takeaways

• MiniMax-H3 (33B, omni-modal, native stereo audio) is the most downloaded video model on Hugging Face: 4.8M total downloads, including ~4.58M over the last 30 days (September 2026).
• LTX-2.5 (Lightricks) surpassed 1.5M downloads one month after its open-weights release (August 11, 2026); the LTX family has accumulated 33M downloads.
• The day-1 ComfyUI integration and community quantizations (pruned INT8 diffusion ~21GB on a single H100) let these models run outside the publishers' datacenters.
• On the text side, Alibaba's Qwen3.8-27B shows ~7.7M downloads: the open-weight shift affects all modalities, not just video.
• Concrete consequence: a professional-quality video pipeline without a subscription becomes realistic — with limitations you need to know about (768p locally, licenses to read).

Recommended Tools

Two strategies are available to you: self-host the open weights, or pay per use via API. Here's the overview.

The Numbers Behind a Measurable Shift

Yes, video generation is becoming a commodity — and for once, no need to take a press release at its word to see it.

The numbers speak for themselves. MiniMax-H3 shows 4.8M total downloads, of which roughly 4.58M came in the last 30 days alone — meaning the bulk of its adoption is recent, following the release of the weights on September 11, 2026. It is the most downloaded video model on the platform over that period, and it sits at the top of the weekly trending chart.

LTX-2.5 follows the same slope. Released as open weights on August 11, 2026, it has surpassed 1.5M downloads in a single month, and the LTX family as a whole crosses 33M cumulative downloads — enough to make it, according to RunPod's deployment guide, "the most widely used line of open world models on the market."

One caveat is in order: a download counter measures adoption, not quality. But adoption is precisely what defines a commodity. When an open-weight model surpasses in usage volume the closed APIs that dominate the benchmarks, the balance of power shifts in nature — the vendor no longer dictates the rules alone.

MiniMax-H3: a single transformer for video and sound

H3 isn't just another text-to-video model: it's a single 33B transformer that generates video along with its stereo soundtrack, where you previously had to stack a T2V model, an I2V model, and a TTS model.

That's the technical difference that explains the buzz. The model's Hugging Face card describes an omni-modal architecture: text, image, video, and audio as input, and video output with native stereo sound. One single transformer, instead of a stack of specialists to chain together — along with the sync inconsistencies that entails.

Concretely, the open weights published on September 11, 2026 cover two variants:

• H3-Base FL2VA: text/image to video + audio, with first and last frame control — exactly what you need for editing workflows;
• Ref2VA: reference-guided generation.

Native generation is 768p. 2K upscaling and Context-IR remain API-side — a deliberate commercial choice, which we'll come back to below. For self-hosting, the community integration documented by Spheron runs the pruned INT8 diffusion model (~21 GB) on a single H100 PCIe, with the Qwen3-VL 32B text encoder pre-quantized in NVFP4/AWQ and separate video and audio VAEs.

My take: native audio is the real story. As long as AI video came out silent, every generation required a separate sound design step. A model that produces image and sound together halves the pipeline — and that's exactly the kind of simplification that turns a technology into a commodity.

LTX-2.5: Speed as a Weapon Against Lock-In

Lightricks isn't trying to win on raw quality. LTX-2.5 wins on rendering time — and the announced numbers are hard to ignore.

The headline stat: 10 seconds of 720p video generated in 6.8 seconds on 2x NVIDIA GB200. Faster than real time. Via the LTX API, a full 1080p clip comes out in 23.7 seconds end-to-end.

An honest caveat: these measurements come from the vendor and are relayed by RunPod. Take them with the usual grain of salt. But the gap is too large to be a measurement artifact.

Under the hood, it's a near-complete rebuild of the pipeline: a diffusion-based video decoder replacing classic VAE decoding (fewer artifacts, fewer melted textures), a custom Gemma 4 12B text encoder with a prompt enhancer, multi-shot generation in a single pass, a duration predictor, and Diffusion Fidelity Rendering — compute allocated based on scene complexity, with a latent compressed 8x in time.

That last point deserves a closer look: allocating compute based on complexity means refusing to pay the maximum price for every frame. It's engineering in service of cost, not of benchmarks.

On pricing, the logic is that of a vendor playing the openness card: free open weights for any organization under $10M in annual revenue, paid licensing above that, and the LTX-2.5 Pro API at $1.20/clip (August 2026) versus ~$1/clip for Gemini Omni Flash and $0.30/s for Veo, according to VentureBeat.

ComfyUI: The Real Infrastructure of Commoditization

What turns open weights into a commodity isn't the model. It's the tool that lets you run it at home, on your own hardware, with no account and no API key.

The day-one partnership between LTX and ComfyUI is the best illustration of this. "By keeping LTX open, we let teams control their hardware, their intellectual property, and their model," summarizes Zeev Farbman, co-founder and CEO of LTX, as quoted by VentureBeat. On the other side, Yoland Yan, co-founder and CEO of ComfyUI, owns the philosophy: "Openness is what lets the community move fast."

And the community does move fast. As soon as MiniMax-H3 was released, the Comfy-Org/MiniMax-H3 integration was already available:

git clone https://github.com/Comfy-Org/MiniMax-H3

Above all, community work doesn't stop at the official weights. Pruned NVFP4 quantizations, a pre-quantized text encoder, and above all a VDN-H3 distillation that brings generation down to 8 steps with near-identical quality: every week, the hardware cost of running it drops. The model that required a cluster at announcement time runs a few weeks later on hardware rented by the hour.

This pattern isn't unique to video. You can see it in code with OpenCode, the open source coding agent outpacing Claude Code and Codex with 8 million developers and 172K GitHub stars, or in inference tooling with the token war and open source tools like Caveman, RTK, and OmniRoute. Everywhere, the same dynamic: labs publish, the community industrializes.

The LLM trajectory, replayed modality by modality

What is happening to video happened to text twelve months earlier, with the same scenario: closed models lock in through quality, open ones catch up through distribution.

The text side of this week's numbers proves it. Qwen3.8-27B, Alibaba's dense vision-language model, posts ~7.7M downloads over the last 30 days — under the Apache 2.0 license, with a native context of 262K extendable to 1M and thinking enabled by default. One symptom, among others, of the same open-weight shift.

The ecosystem scale is even more telling. Alibaba claims the most downloaded open-source family in the world: over 460 published models, over 300,000 derivative models, and over 3 billion cumulative downloads. Qwen derivatives alone account for 151,448 repositories on the Hub — 2.6 times Meta's footprint, according to Alibaba Cloud.

The pattern had already been seen in other domains: Cohere North Small Translate, an open-weight translation MoE that beats DeepL and Google Translate on WMT26, or Moonshot AI's Kimi K2.6, which dominates open-weight while China raises $2 billion to accelerate. Text, translation, code, and now video: every modality follows the same S-curve.

My read: closed APIs won't disappear — they will remain the top end of the quality range. But the floor is rising, and it's the floor that defines what a creator can afford. When the floor is free, the leverage of monthly subscriptions collapses.

What this actually changes for creators

A professional-quality video pipeline without a subscription is no longer a fantasy: it's an informed architectural decision.

The hybrid strategy is today the most rational one:

1. Iterate locally with quantized open weights (free, 768p, distilled to 8 steps) — you can break 50 prompts without keeping count.
2. Render via API only the final shots that require 2K or Context-IR: $1.20/clip at LTX Pro, ~$1/clip at Gemini Omni Flash, $0.30/s at Veo (August 2026).
3. Own your pipeline: reproducible ComfyUI workflows, no unilateral pricing changes, no model pulled overnight.

The math is telling. A hundred prompt iterations locally cost the price of a few hours of GPU rental. The same hundred iterations on a closed API at $0.30/s on 10-second clips: $300. Massive iteration no longer needs a subscription — it needs a graphics card, even a rented one.

To distribute the result, there's no need to depend on a platform that can demonetize your videos overnight: your own site, hosted for example on Hostinger, is enough to own your distribution. And if you're assembling the full chain — video, music, editing — our selection of AI tools for video editing and best free AI tools completes the pipeline without spending a euro.

The rest of the creative ecosystem, for its part, is following the same professionalization: Suno v6 arrives with Warner BMG and Believe and licensed-trained music models, a sign that creative generation is leaving the gadget era and entering that of production tools.

The limits remain, and they are real: 768p locally, hardware to plan for, licenses to read region by region. Convenience isn't universal free access — it's an accessible floor.

❌ Common Mistakes

Mistake 1: Believing that "open weights" means "free with no strings attached"

The LTX license is free under $10M in annual revenue, paid beyond that. Self-hosting MiniMax-H3 is subject to licensing in certain regions. The solution: read the license before building a business model on top of it, not after.

Mistake 2: Comparing API demos to local weights

The demos circulating are often in 2K via the API, with upscaling and Context-IR. The open weights generate 768p at base. The solution: calibrate your expectations to what you actually run, and keep the API for the final render.

Mistake 3: Underestimating the required hardware

MiniMax-H3's pruned INT8 diffusion weighs ~21 GB and targets an H100 PCIe, not counting the Qwen3-VL 32B text encoder and the video and audio VAEs. The solution: start with the NVFP4/AWQ quantizations and the VDN-H3 8-step distillation, or rent a GPU by the hour — the Spheron and RunPod guides cover deployment step by step.

Mistake 4: Taking download counts as a quality verdict

4.8M downloads measure adoption and curiosity, not technical superiority. Closed APIs retain the lead on certain quality criteria. The solution: test on your own use cases — your style, your durations, your editing constraints.

❓ Frequently Asked Questions

Does MiniMax-H3 really generate video with sound?

Yes. That's even its defining feature: a single 33B transformer produces the video with native stereo audio, in an omni-modal architecture (text, image, video, and audio as input). In the ComfyUI integration, the video and audio VAEs are separate, but generation remains unified — no need to chain a TTS model afterward.

What GPU do you need to run MiniMax-H3 locally?

The community benchmark is the pruned INT8 diffusion (~21 GB) on a single H100 PCIe. NVFP4 and AWQ quantizations of the Qwen3-VL 32B text encoder further reduce the footprint, and VDN-H3 distillation at 8 steps significantly speeds up generation. Without a GPU, the Spheron and RunPod guides detail hourly-billed cloud deployment.

Can LTX-2.5 be used commercially?

Yes, with conditions: the license is free for any organization making less than $10M in annual revenue. Above that, a paid license applies. The LTX-2.5 Pro API is billed at $1.20/clip (as of August 2026, check Lightricks' website). For the majority of independent creators and SMEs, commercial use of the open weights therefore remains free.

Why are the open weights only released at 768p?

Because the 2K upscale and Context-IR remain API-side at MiniMax: it's the monetization that funds the open model. The strategy is classic — free at the entry level, paid in high definition. Community upscalers close part of the gap, without matching a native 2K pipeline.

Should you abandon Veo, Kling, and closed APIs?

No. Closed models keep the quality advantage and the simplicity — Veo at $0.30/s remains the go-to for productions that buy ready-made quality. But they are no longer the only serious option. The right question isn't "which one is best," but "which one at which stage of the pipeline": open for iterating, closed for finalizing if needed.

What's the connection between Qwen3.8-27B and video generation?

None directly — it's a vision-language model, not a video generator. But its ~7.7M downloads over 30 days show that the open-weight shift is hitting all modalities at the same time. Video is not an exception: it's the last chapter of the same story, the one where open weights became the ecosystem's default.

✅ Conclusion

With 4.8M downloads for MiniMax-H3, 1.5M for LTX-2.5, and 7.7M for Qwen3.8-27B, open-weight video generation has crossed the convenience threshold — and your creative pipeline no longer needs a subscription to be professional. To make sure you don't miss what comes next, follow our watch of new AI tools.