Skill LLM MiniMax China Mainland
Purpose
Build a validated MiniMax China Mainland request for a model already selected by the user or host project. This skill is not a model selector or execution layer.
Read Order
Read shared schemas first, then:
references/connection-profiles.mdreferences/request-urls.mdreferences/model-catalog.mdreferences/capability-matrix.mdreferences/model-parameters.mdreferences/api-surface-scope.md- the matching transport, including
references/transport-video.md for MiniMax-H3
references/pricing-matrix.md only when cost mattersreferences/model-sync.md only during an explicit update
Runtime Rules
- Provider identifier is
minimax-cn; maintained IDs are exactly MiniMax-M3, MiniMax-H3, image-01, and music-3.0.
- China Mainland and International endpoints, credentials, currencies, and prices are not interchangeable.
- Resolve the full route key before using any parameter.
- MiniMax-M3 OpenAI compatibility defaults thinking on; Anthropic compatibility defaults thinking off. Never normalize those defaults into one provider-wide rule.
reasoning_split changes response layout only; it does not enable or disable thinking.- Prefer
max_completion_tokens; max_tokens is legacy on OpenAI compatibility.
- Fields ignored by the Anthropic compatibility layer are not supported controls. Reject them rather than assuming they work.
- Image, video, and music payloads are native MiniMax contracts, not OpenAI/Anthropic messages.
MiniMax-H3 uses the regional root domain and native asynchronous V2 routes; never append H3 paths to a /v1 media base URL.- H3 primary generation, H3-Context-IR, and H3 regeneration are distinct surfaces with distinct required fields and result mappings.
- H3 requires MiniMax pay-as-you-go API access. Do not infer a special key type, and do not substitute a legacy video model, package, API surface, or region.
unknown is fail-closed. Do not infer support or absence from a generic schema or a single-model example.
Request Kinds
chat: text-first generation and toolsvision: image/video understandingimaging: text-to-image or subject-reference image-to-imagevideo: MiniMax-H3 primary generation, H3-Context-IR prompt enhancement, or 768P-to-2K regenerationmusic: text/lyrics-to-music
Out of Scope
Other MiniMax models, arbitrary-video processing, regional fallback, model ranking, automatic substitution, request execution, and implicit parameter translation between surfaces.
1---2name: skill-llm-minimax-cn3description: Runtime contract for MiniMax China Mainland direct-model integration. Use it to resolve exact retained model IDs, regional endpoints, OpenAI-compatible and Anthropic-compatible text surfaces, per-surface thinking defaults, multimodal input, tools, MiniMax-H3 video workflows, image generation/editing, Music 3.0, streaming, pricing, and response mapping. It constructs requests only and never executes them.4---56# Skill LLM MiniMax China Mainland78## Purpose910Build a validated MiniMax China Mainland request for a model already selected by the user or host project. This skill is not a model selector or execution layer.1112## Read Order1314Read shared schemas first, then:15161. `references/connection-profiles.md`172. `references/request-urls.md`183. `references/model-catalog.md`194. `references/capability-matrix.md`205. `references/model-parameters.md`216. `references/api-surface-scope.md`227. the matching transport, including `references/transport-video.md` for `MiniMax-H3`238. `references/pricing-matrix.md` only when cost matters249. `references/model-sync.md` only during an explicit update2526## Runtime Rules2728- Provider identifier is `minimax-cn`; maintained IDs are exactly `MiniMax-M3`, `MiniMax-H3`, `image-01`, and `music-3.0`.29- China Mainland and International endpoints, credentials, currencies, and prices are not interchangeable.30- Resolve the full route key before using any parameter.31- MiniMax-M3 OpenAI compatibility defaults thinking on; Anthropic compatibility defaults thinking off. Never normalize those defaults into one provider-wide rule.32- `reasoning_split` changes response layout only; it does not enable or disable thinking.33- Prefer `max_completion_tokens`; `max_tokens` is legacy on OpenAI compatibility.34- Fields ignored by the Anthropic compatibility layer are not supported controls. Reject them rather than assuming they work.35- Image, video, and music payloads are native MiniMax contracts, not OpenAI/Anthropic messages.36- `MiniMax-H3` uses the regional root domain and native asynchronous V2 routes; never append H3 paths to a `/v1` media base URL.37- H3 primary generation, H3-Context-IR, and H3 regeneration are distinct surfaces with distinct required fields and result mappings.38- H3 requires MiniMax pay-as-you-go API access. Do not infer a special key type, and do not substitute a legacy video model, package, API surface, or region.39- `unknown` is fail-closed. Do not infer support or absence from a generic schema or a single-model example.4041## Request Kinds4243- `chat`: text-first generation and tools44- `vision`: image/video understanding45- `imaging`: text-to-image or subject-reference image-to-image46- `video`: MiniMax-H3 primary generation, H3-Context-IR prompt enhancement, or 768P-to-2K regeneration47- `music`: text/lyrics-to-music4849## Out of Scope5051Other MiniMax models, arbitrary-video processing, regional fallback, model ranking, automatic substitution, request execution, and implicit parameter translation between surfaces.
joeylu