Models

Clients ask for a model group, a stable name like deepseek-flash. The gateway maps the group to a provider, applies fallbacks, and meters the call. A provider change does not change client code.

Model groups

deepseek-v4-pro falls back to deepseek-flash when the reasoning model is unavailable. DeepSeek thinking is on by default for both groups (see Chat completions). Video groups are asynchronous task APIs; see Video generation and BytePlus Seedance.

A deployment serves only groups whose provider credential is configured. GET /v1/models is the authoritative list for a key and deployment; this table describes the gateway's full catalog.

deepseek-flash

DeepSeek Flash, the default. Text and image input. Thinking is enabled by default and reasoning_effort controls it. Alias: deepseek-chat.

deepseek-v4-pro

DeepSeek Pro, tuned for reasoning. Thinking is enabled by default; reasoning_effort accepts none, low, high (default) or max. Alias: deepseek-reasoner.

minimax-h3

MiniMax H3 on /v1/videos at 768P or 2K, 4–15 s.

minimax-h3-max

MiniMax H3-Max on /v1/videos at 480P or 768P, 5–15 s.

seedance

BytePlus Seedance 2.5 on /v1/videos at 480p, 720p or 1080p, 4–30 s, from text or one first-frame image. See BytePlus Seedance for the full capability and pricing notes.

Check what your key can call

A key only sees the groups it was issued for. If a group is missing, ask your administrator for a key that includes it.

Choosing a group

• Use deepseek-flash for fast text and image input.
• Use deepseek-v4-pro for multi-step reasoning, planning and code review.
• Use minimax-h3 or minimax-h3-max for clips of 4–15 s.
• Use seedance for clips up to 30 s, including image-to-video.