Self-hosted H3 video server: API key, TLS and read-only media

Before you put a self-hosted MiniMax H3 server on a public address, MiniMax's guide asks for three things: API key authentication, TLS termination, and read-only mounts for any server-local reference media, limited to dedicated directories. The SGLang server it describes listens on a plain HTTP port, and the commands shown do not set up authentication or TLS, so the front door is your job.

What does the vendor guide require?

These items come from the "Key Constraints" part of the self-hosting guide (read 2026-10-09). The wording below is a paraphrase.

• Require API key authentication before public exposure.
• Terminate TLS in front of the server.
• Mount server-local reference media read-only.
• Allow only dedicated directories for reference files, not the whole disk.
• The guide's SGLang example binds 0.0.0.0; change that to 127.0.0.1 unless a proxy with those controls sits in front.

Why do reference files matter?

The Ref2VA variant takes up to 9 images, 3 video clips and 3 audio clips. If a request can name a path on the server, a careless setup lets a caller point the model at files that should not be read. Restricting the readable directories is the guide's answer.

What would a minimal setup look like?

Bind the video server to 127.0.0.1 (the guide's example uses 0.0.0.0, so change it) and put a reverse proxy in front that checks a key and terminates TLS. Mount the reference directory read-only into the container or service, and give the server only that one directory. Log each request with the caller and the prompt length so a spike is easy to trace.

Add a job queue ahead of the GPU. One H3 job already occupies a whole node of GPUs on the reference setup, so a public endpoint without a queue lets one client block everyone else. This part is not from the vendor's guide; it follows from the load, and you should size it yourself.

What does a hosted API already do?

On Sume the call is authenticated with Authorization: Bearer $SUME_API_KEY. A callback_url must be HTTPS, and Sume signs the raw JSON body and sends x-sume-webhook-timestamp and x-sume-webhook-signature headers, so your receiver can verify who sent it (see the webhooks guide). Reference media is sent as public HTTPS URLs, not server paths.

One cost of hosting: Sume has no Zero Data Retention toggle (see the privacy docs). If that is a hard requirement, it is a reason to self-host, and a reason to read the license too: it excludes the US, EU, UK and South Korea from the territory, and it asks for written authorization above 20 million USD yearly revenue.