Video production is moving from "by appointment" to "on demand."
The biggest shift in video generation this year isn't a new SOTA—it's open weights. Top-tier results are now accessible on any team's desktop. MiniMax-H3 and LTX-2.5 are two leading examples: both open-source, both often compared in the community to commercial flagships like Seedance 2.0. The key difference? Weights are open, local deployment is supported, and you can modify them freely.
This week, we ran a local evaluation designed to mirror real production. We fed the same prompts to both models on the same M114 blade, across four durations: 5s, 10s, 15s, and 30s. The results were more interesting than expected.
I. The Setup
To keep things reproducible, we held variables constant:
- Models: MiniMax-H3 and LTX-2.5, both running locally with open weights.
- Prompts: Identical prompt set for each model.
- Durations: 5s, 10s, 15s, 30s—run sequentially.
- Hardware: Single node on the E1005 Pro desktop intelligent computing supernode.
II. Results
Bottom line: All eight combinations completed successfully on the same M114 blade. No OOM, no timeouts.
Focusing on the heaviest case (30s), which stresses hardware the most:
- MiniMax-H3: 30s 480P video in ~6,000 seconds (~100 min).
- LTX-2.5: 30s 720P video in ~1,000 seconds (~17 min).
(Video playback failed. Please refresh and try again.)
The videos above use the same prompts, so you can directly compare image quality and motion handling. (Prompts are in the comments—feel free to grab them.)
III. Where They Diverge
Both models natively support audio-video sync—soundtracks are generated out of the box, no separate dubbing or alignment needed. That's a big win for local deployment: one less import/export step.
The main difference? Language control. MiniMax-H3 lets you set dubbing language via prompts (we used Chinese in this test). LTX-2.5 defaults to English. If your primary output is Chinese, that's a decision point upfront.
Another similarity: neither requires elaborate prompt engineering—short descriptions work. The real gap is comprehension.
- MiniMax-H3 excels at understanding. It accurately interprets camera moves, actions, and atmosphere—delivering visuals that match your intent. For high-quality projects with costly rework, this reduces miscommunication and re-dos.
- LTX-2.5 wins on openness. It exposes not just weights but the training framework, allowing deep customization and seamless integration into your existing toolchain. If you need to embed generation into your workflow and adapt it to specific business needs, LTX-2.5 gives you more room.
Neither model is "better"—they're positioned differently: fidelity and fewer reworks? MiniMax-H3. Customization and scalability? LTX-2.5.
IV. The Real Takeaway: Hardware Shouldn't Lock You In
What's more important than "which model is best"? That all eight combinations ran on the same blade.
Model choice shouldn't be a commitment. This month you might need MiniMax-H3 for fine-grained control; next month, LTX-2.5 for rapid scaling; six months later, a new open model. Durations vary too—short clips for social, 30s+ for promos.
Hardware is a one-time investment—you can't swap it out every time you switch models. So the real value of a local device isn't how well it runs any single model, but that it handles whatever you throw at it—regardless of model or duration.
Four durations, two models, eight combinations—all on one M114 blade. That's what we wanted to validate.
The choice stays with you, not with hardware limitations. That's the confidence local deployment should deliver.
V. Local vs. Cloud: How to Choose
For content teams, this isn't a technical debate—it's about cost and security.
- Unlimited iteration. Cloud charges per clip or per second—every version costs money. Local is a one-time investment; marginal cost per extra run is near zero (just electricity). If you generate high volumes to filter for the best ideas, you can iterate freely.
- Data stays on-prem. Footage is your lifeblood. Cloud generation means assets leave your network. Local deployment keeps data on your machine—no risk of theft. For brands and film studios, this often matters more than speed.
- You control the pace. Cloud has queues, rate limits, and service changes. Local runs offline—you're in full control.
VI. The Storage-Compute Foundation: E1005 Pro Desktop Supernode
This evaluation used a single M114 node in the E1005 Pro, with: 14-core CPU, 128GB shared memory, 8TB storage, 10GbE networking, expandable to 245TB storage.
(Image: M114 single node left; E1005 Pro full system right)
The E1005 Pro supports up to 5 M114 nodes in a clustered deployment, all in one appliance about the size of a coffee machine—plug and play. No server room, no power grid modifications.
Desktop AI devices aren't rare, and 128GB-equivalent VRAM isn't unheard of. What sets the E1005 Pro apart is what happens after the model starts running.
Three advantages:
- Compute scales horizontally, staying a single system. On standalone boxes, VRAM and compute are fixed—if you hit limits, you buy another machine and manage two incompatible systems. With the E1005 Pro, adding a blade adds compute, and with five blades it's still one system, one namespace. As you grow, the infrastructure grows with you.
- Multiple pipelines run simultaneously. When a single blade is busy, others wait. In a multi-blade config, text-to-image, text-to-video, and LLMs can run in parallel—storyboards, visuals, and asset management happen concurrently. That speeds up the entire delivery.
- Storage built for media, not just the OS. Desktop devices often have 1–4TB SSD, which fills quickly. A single M114 node expands to 245TB—project files, asset libraries, and exports stay on the same machine. Assets don't move, so data leakage risk is minimal.
The question isn't just "can it run a big model"—it's "can it grow with your business."
VII. From Running a Model to Producing a Video—There's a Platform in Between
Running a model is just step one. Any video maker knows production is a pipeline: script, storyboard, character/scene consistency, cinematography, rendering, editing—all require human input. The model solves only one piece.
That's why the E1005 Pro integrates a full AI-powered video creation suite into the device: from script to storyboard, visuals to 3D virtual studio, to multi-model rendering—all in one platform.
Three layers of impact:
- Democratizes pro workflows. Breaking down storyboards, shot sequencing, character consistency—once learned through experience—are now built-in steps. You write an idea, the platform guides you through.
- Gives professionals time back. Characters, scenes, props are reusable across shots. The 3D virtual studio lets you preview blocking and camera moves before generation. The storyboard stage includes a "production gate" that catches issues (missing dialogue, wrong sequence, incorrect duration) before you spend compute. Every avoided failure saves real money and machine time.
- Turns "gacha" into directing. Structured shot descriptions are auto-translated into prompts for different models, letting creators choose the best model—not gamble.
The model sets the ceiling for a single shot; the platform determines whether the film delivers on time.
VIII. Final Thoughts: What Problem Are We Actually Solving?
In conversations with content teams, studios, and brands, the pain points consistently reduce to three:
- Money wasted on trial and error. Success rates are low—behind every usable video are many failed generations. Cloud charges per run, so your budget pays for failures. With local deployment, marginal iteration cost is near zero—those dozen tries become part of the creative process, not a write-off.
- Assets that can't leave the premises. Unreleased materials, scripts, character designs, client data—once in the cloud, you lose control. "Data stays on-prem" isn't a slogan; it's a prerequisite for AI adoption.
- Workflows that break halfway. The model runs, but all the steps between script and final export are still manual. That's why we built a unified platform—so non-experts can produce, and experts can skip repetitive labor.
These three issues drove us to build the E1005 Pro and software suite: storage-compute, models, and the creative workflow, all in a coffee-machine-sized box. No server room, no dedicated staff—just turn ideas into finished videos.
This evaluation was simple: four durations, two models, eight combinations, all on one blade. But the message isn't: great models will keep emerging, and the infrastructure's job is to let you choose freely every time—not settle.
Models change, durations vary, but creativity shouldn't wait for a scheduled slot. If you're evaluating on-prem solutions for video generation, we'd love to hear how you're using it.
