MiniMax H3 open weights vs a hosted lip-sync API: what you take on

If you want a talking clip from a photo and a voice line without running GPUs, use a hosted route; if you want to run MiniMax H3 yourself, the open weights are a separate starting point, and nothing in them is a ready-made lip-sync service. A tracker lists MiniMax H3 as released on 2026-07-31 with open weights on 2026-08-03 (Magic Hour tracker, read 2026-10-06). The same page describes the model as 15 seconds, 2K and native stereo audio. It does not describe a lip-sync product.

On Sume, H3 appears twice and neither is self-hosting. The video catalog lists minimax-h3 and minimax-h3-max for prompt-driven video, and the MiniMax H3 Max Lip Sync route, minimax/h3-max/lip-sync, takes a still plus audio of 5 to 14.8 seconds (Models and Video generation, read 2026-10-06). Sume does not sell or ship weights.

What do you own if you self-host?

When does self-hosting make sense?

Self-hosting makes sense when volume is high and steady enough to keep hardware busy, when data must stay inside your network, or when you need to modify the model. It is a bad fit for a team that wants ten talking clips this week, because the setup and tuning cost arrives before the first clip does. A hosted route also changes who tracks provider updates: Sume's docs keep the provider queue id internal and expose only the public model id.

Two cautions apply to either path. Open weights are not automatically licensed for the use you intend, so read the release terms before you ship ads built on them. And a video model that generates its own audio is not a lip-sync tool: Sume's own guidance is that a talking face is a still plus audio through a lip-sync model, never a video-model clip with narration under it.

How do you try the hosted route in an afternoon?

Generate or pick a still, produce a voice line of 5 to 14.8 seconds with Sume TTS, and submit both with duration_seconds set to the measured length and resolution: 480p for a cheap first look. If the lips and teeth hold up, raise it to 768p. Each job reserves its cost at admit, and failed jobs are refunded. Because Sume accepts valid jobs as queued when your workspace is at its concurrency limit, as long as queue capacity remains (a Pro workspace processes 4 at a time with a default queue of 20), you can submit a batch of variants at once and read the results as they finish (Generation admission).

The fresh-audio post shows how to get audio that fits the window, and the MiniMax price table lists every MiniMax id on Sume side by side. If a test shows you need more control than a hosted route gives, that is your evidence to go self-hosted, and you will have a baseline clip to compare against.

What would self-hosting change for you?

Open weights move cost and operations onto you. You would provision GPUs, keep a runtime current, handle queueing and decide what happens when a job fails halfway. Sume's hosted route replaces that with a per-second price, a reservation taken at admit, a refund on failure and a plan-based concurrency limit. Neither is free; they trade a bill you can predict for a system you must run.

This page does not claim the weights run on Sume or that a hosted route behaves the same as a self-hosted checkpoint. It only says the hosted H3 Max lip-sync route exists, with documented limits. If you need the weights for a reason such as data residency, evaluate them on your own infrastructure and keep the hosted route for the work where a per-second invoice is cheaper than a cluster.