LTX 2.5 vs MiniMax H3: Specs, Real Costs, Hardware

LTX 2.5 vs MiniMax H3 is a choice between an open-weights model you host yourself and a 2K model that runs the moment you type a prompt — and the gap between them is hardware, not quality. LTX-2.5 shipped as open weights on August 11, 2026; MiniMax H3 already runs on Kavel at a fixed 2K output. Here is what each one costs and what each one needs.

LTX 2.5 vs MiniMax H3, short answer: LTX-2.5 is the more capable model — native multi-shot, synchronized audio, up to 4K, up to 20 seconds — and its weights are free to use commercially under $10M revenue. But self-hosting it wants 32GB+ VRAM minimum, A100 80GB or H100 recommended, per Lightricks' own system requirements. Through LTX's managed API it costs $0.19/sec at 1440p. MiniMax H3 costs $0.105/sec at a fixed 2K — 45% less per second at a comparable resolution — with nothing to install. Pick LTX-2.5 if you need 4K, multi-shot or audio and have the GPU (or the API budget). Pick H3 if you want 2K clips now. Verified August 12, 2026 against official docs.

LTX 2.5 vs MiniMax H3: the spec sheet, side by side

LTX-2.5 wins the row count outright, and it is worth being precise about why: the 2.5 release added a Diffusion Video Decoder, native multi-shot scenes that hold character, lighting and voice across cuts, and an automatic-duration mode where you send "duration": null and the model picks the length from your prompt. H3 does none of those four.

The question is what it takes to get them — which is where the comparison stops being about the models.

The 32GB line nobody quotes

Social summaries of the launch circulated a figure of roughly 16GB VRAM to run LTX-2.5 locally. Lightricks' own system requirements page says otherwise:

GPU: NVIDIA GPU with a minimum 32GB+ VRAM - more is better
RAM: 32GB system memory
Storage: 100GB free space

And the recommended configuration is an NVIDIA A100 (80GB) or H100 — data-center parts, not desktop cards. Community int8 and GGUF quantizations do bring the footprint down, and people are genuinely running them; that is a lossy variant of the model, though, not the 22B checkpoint the specs above describe. Both things are true at once, and only one of them is in the official docs.

The launch's headline speed number carries the same footnote. LTX-2.5 generating a 10-second 720p clip in 6.8 seconds is real — measured self-hosted on two NVIDIA GB200 chips at steady state. Through LTX's own managed API, the comparable job reportedly took 23.7 seconds at 1080p. "Faster than real time" is a statement about a GB200 rack, not about your machine.

LTX 2.5 vs MiniMax H3 pricing: what a clip actually costs

LTX's published per-second API rates, read from docs.ltx.io/pricing on August 12, 2026:

Two things that table tells you. The Pro variant is capped at 1080p — for 1440p or 4K you must use Fast, so "pro quality at 4K" is not an option that exists. And the ladder climbs steeply: 4K costs 3.3× what 720p does, per second. H3 has no equivalent decision to get wrong.

H3 bills one flat rate

MiniMax H3 has no ladder at all. The H3 API accepts one output setting — 2K — at $0.105 per second, flat. On Kavel that works out to 31.5 credits per second: a 5-second H3 clip is 158 credits, a 10-second one is 315.

At the nearest matching resolution:

H3 is 45% cheaper per second in that band. LTX-2.5 undercuts H3 at 720p ($0.09/sec) — but that is a lower resolution being compared to H3's 2K, which is not the same deliverable.

LTX 2.5 vs MiniMax H3: what H3 actually produces

These were generated on H3 through Kavel — a product turntable, an FPV chase, and a held portrait composition, covering what people usually test a video model on:

A luxury perfume bottle rotating on a lit pedestal inside a gold wireframe cube — product-in-motion on MiniMax H3.

A drone-style FPV chase following a motocross rider through a ruined street — camera movement on MiniMax H3.

H3 has no quality selector, and that is the point

H3's settings surface is deliberately short: pick text-to-video or image-to-video, write the prompt, choose 5, 6, 8 or 10 seconds. There is no quality dropdown on H3 because there is nothing to select — every H3 generation comes out at 2K, so the resolution decision that costs you three attempts on a laddered model simply does not exist here. For a fuller walkthrough of the model and where its name came from, see our guide to Hailuo 3.0 / H3, and the MiniMax H3 generator itself.

LTX 2.5 vs MiniMax H3: which one for which job

Reach for LTX-2.5 when the deliverable needs something H3 structurally cannot do. A 20-second take. A sequence that cuts between shots and keeps the same character. Synchronized audio baked into the generation. True 4K for a screen larger than a laptop. Those are not preferences; they are hard capability lines, and LTX-2.5 is on the right side of all four.

Reach for MiniMax H3 when the job is a 2K clip and the constraint is time or money. With H3 there is no CUDA version to match, no 100GB download, no A100 to rent. At $0.105/sec it is also the cheaper of the two in the resolution band where they overlap.

And the two are not mutually exclusive. The open-weights license is genuinely permissive — free for commercial and production use under $10M in annual revenue, with a paid agreement above that — so if you have the hardware, LTX-2.5 as a self-hosted workhorse plus H3 for everything quick is a coherent stack.

Want to see what 2K looks like on your own prompt? Run it on MiniMax H3 — H3 credit costs, and every other model's, are listed on the pricing page.

FAQ

Is LTX-2.5 free?

The weights are free to download and free for commercial and production use if your organization is under $10M in annual revenue; above that, Lightricks requires a paid commercial agreement. Running it is not free — you supply a 32GB+ VRAM GPU, or you pay LTX's API rates starting at $0.09/sec at 720p.

How much VRAM do you need for LTX-2.5?

Lightricks' published minimum is a 32GB+ VRAM NVIDIA GPU with 32GB of system RAM and 100GB of free storage; the recommended configuration is an A100 80GB or H100. Quantized community builds run in less, at some quality cost.

Is MiniMax H3 the same thing as Hailuo 3.0?

Yes — H3 is MiniMax's Hailuo 3.0 line, and it is served under the identifier hailuo-03. Both names refer to the same model.

Which is cheaper, LTX-2.5 or MiniMax H3?

At comparable resolution, H3: $0.105/sec at 2K versus $0.19/sec for LTX-2.5 at 1440p through the LTX API. LTX-2.5 is cheaper than H3 only if you drop to 720p ($0.09/sec) or self-host on hardware you already own.

Can MiniMax H3 generate 4K or multi-shot video?

No. H3 outputs a fixed 2K and one continuous shot per generation, up to 10 seconds. Native multi-shot and 4K are LTX-2.5 capabilities, not H3 ones.

Does LTX-2.5 generate audio?

Yes. LTX-2.5 produces synchronized video and audio, and it accepts audio as an input mode as well; you can disable it with generate_audio: false for silent output. The H3 generator does not expose audio, so plan on scoring those clips yourself.

Sources

Every LTX-2.5 spec and rate above is read from Lightricks' own documentation on August 12, 2026. The H3 figures come from its provider's published rate card and from Kavel's live H3 configuration. The one third-party number — the 6.8-second benchmark — is attributed below.

• LTX-2.5 model documentation — support matrix, endpoints, automatic duration
• LTX API pricing — per-second rates by model and resolution
• LTX open-source system requirements — VRAM, RAM, storage, CUDA
• Lightricks/LTX-2.5 on Hugging Face — 22B parameters, license terms, distilled variant
• LTX.io launch announcement — August 11, 2026
• VentureBeat: LTX-2.5 speed benchmark — the 6.8s / GB200 measurement and the 23.7s API comparison
• MiniMax H3 on Kavel — the H3 rate, fixed 2K output and 5–10s duration options, as configured live

Last updated: August 2026