FastH3 vs H3 Max: Open Weights or a Browser Clip?

Choose FastH3 when owning the weights, runtime, and deployment is the point. Choose a hosted H3 Max route when you want to submit a shot and receive a playable file without building the inference stack. Neither choice is simply “the faster H3.” They solve different operational problems, and the names now overlap enough to make a casual comparison misleading.

There is one more detail for this site: ordinary text and image jobs in AI Live Generator use fal's H3 Max Turbo, not FastH3 and not MiniMax's direct MiniMax-H3-Max API model. The generator shows its own credit cost before submission.

First, untangle the names

Four labels are circulating around the same model family:

“H3 Fast” is also used by some third-party products as a plan or speed label. It does not automatically mean FastH3. Before comparing a demo, invoice, or benchmark, copy the exact checkpoint or endpoint ID. The adjective on a pricing card is not enough.

FastH3 itself is a family rather than one immutable build. The FastVideo repository describes the original Preview as an open-weight, four-forward sparse distillation of MiniMax H3. It later added an eight-forward V2, an experimental smaller Trim checkpoint, local Apple Silicon support, and reference-conditioned OmniRef releases. A result for one checkpoint on one machine should not be presented as a permanent number for all FastH3 versions.

The real difference is ownership

With FastH3, the model is only the beginning. You also own the installation, checkpoint choice, quantization, memory strategy, queue, file storage, monitoring, and upgrades. That can be exactly what a research group or infrastructure team wants. It also means a creative team cannot judge the product by download size alone.

The current FastH3 OmniRef inference guide makes the trade-off concrete. Its documented local profiles include RTX 5090, RTX PRO 6000, DGX Spark, and an RTX 4090 path with more aggressive loading and streaming. Apple Silicon uses a separate MLX route. Those profiles are engineering recipes, not a browser compatibility list.

A hosted H3 Max route moves that machinery behind an API. You choose the supported inputs, duration, resolution, and controls; the provider handles the model server. That is less flexible at the systems layer and far simpler at the shot layer. You pay for requests instead of turning GPU utilization into a product of your own.

Calling one side “free” and the other “paid” hides the useful part. Public weights let you replace a provider's per-request compute invoice with your own infrastructure bill; they do not make generation costless.

FastH3 has more than text-to-video now

The first FastH3 Preview is easy to summarize: a distilled base-H3 text-to-video-with-audio checkpoint. OmniRef is a different contract. Its current documentation accepts a prompt plus up to nine images, three videos, and three audio clips, with no more than 12 references total. It outputs video with synchronized stereo audio at 24 fps and documents 5, 10, and 15-second lengths.

That makes OmniRef relevant to teams building a custom reference pipeline. It does not make every FastH3 checkpoint reference-aware, and it does not turn the ordinary generator on this site into an OmniRef frontend.

Hosted H3 Max is also not one universal request body. MiniMax's current V2 schema lists MiniMax-H3-Max alongside H3, H3-Highspeed, and H3-Lite. fal separates regular text/image routes from Reference-to-Video, Recast, Lip Sync, Camera Controls, and other specialist jobs. The Max versus Turbo guide maps the routes used by this product.

The practical rule is simple: compare the input contract you actually need. “Supports references” is too vague if your job requires a voice file, two identity photos, a last frame, or a source video.

Cost only becomes comparable after you choose the boundary

MiniMax's current pricing overview lists direct MiniMax-H3-Max output at $0.05 per second for 480P and $0.08 per second for 768P. fal's standard H3 Max rates match those output prices before its current promotion is applied; H3 Max Turbo has lower published standard rates at $0.025 and $0.04 per second. fal is showing a 35% promotion through October 31, 2026, so a discounted screenshot should not be used as a November budget.

Those are provider API prices, not this site's retail credit values. A regular generation here costs 3 credits per second at 480P or 5 credits per second at 768P. The shortest five-second clip is 15 or 25 credits. Plans and packs change the dollar value of a credit, so use the total shown on the Generate button or the pricing page.

FastH3 needs a different calculation:

local cost per accepted clip =
  (GPU rental or amortized hardware + energy + storage + operator time)
  / accepted clips

hosted cost per accepted clip =
  billable request cost / accepted clips

Use the same definition of “accepted” on both sides. A cheap attempt that misses the spoken line or changes the subject is not a cheap deliverable. For local work, include model loading and compilation when they affect the user's wait; for hosted work, include queueing and file delivery. Comparing a bare inference timer with click-to-play latency produces a flattering number, not a useful decision.

A fair test brief

Do not start with two unrelated showcase videos. Write one shot and its rejection rules before either model runs. This original ten-second brief is designed to expose timing, object permanence, camera discipline, and audio:

Ten-second continuous shot inside a quiet neighborhood bicycle workshop at sunrise. A mechanic in a rust-red apron rolls a loose front wheel across the workbench, catches it with one hand, then says, “That wobble is finally gone.” Medium-wide 35mm framing. The camera makes one slow lateral move to the right; no cut and no zoom. Warm window light, pale blue shadows. Audio: the wheel hum fades as the hand catches it, one small metal click, clear dry dialogue, no music. Keep the wheel round, the spokes distinct, and both hands anatomically consistent.

Score the complete file, not its best frame:

1. The wheel rolls, decelerates, and stops in the correct order.
2. The catch happens before the line begins.
3. The wheel stays circular and the spokes do not dissolve.
4. The mouth stops moving when the sentence ends.
5. The camera completes one lateral move without a hidden cut or push-in.
6. The hum and click land on their visible causes.

Keep prompt, length, resolution, aspect ratio, and reference assets matched where both workflows allow it. Record every difference that cannot be matched. This is a test plan, not a claim that we ran the benchmark or that one model always wins it.

Which one should you choose?

Pick FastH3 if you need a public checkpoint, want to tune the serving path, must keep generation inside infrastructure you control, or are prepared to make GPU utilization part of the product. Read the exact model card and MiniMax H3 Community License before commercial use or redistribution; “open-weight” is not shorthand for unrestricted use.

Pick a hosted H3 Max route if the job is to make and ship clips rather than maintain an inference service. Then choose by exact route: Turbo for the ordinary text/image workflow used here, standard Max when a provider-specific reason justifies it, and a specialist endpoint when its required inputs match the assignment.

If you are comparing H3 more broadly, MiniMax H3 versus H3 Max explains the base-model split, and H3 Max limitations lists what the browser workflow here does and does not expose.

The cleanest answer is not “FastH3 is faster” or “Max is better.” FastH3 gives you a model stack to own. H3 Max gives you a service contract to use. Choose the burden you are actually equipped to carry.