Upstream support matrix
Lumoswitch currently includes 33 inference-channel presets. Each preset keeps the stable channel code, upstream wire protocol, and LiteLLM provider separate. It also supplies the recommended base URL and model-discovery rules. Channel permissions determine which models are visible; structured capability declarations come from the catalog and administrator overrides, never guesses from /models.
The compiler uses the explicit, reviewed LiteLLM provider prefix; it never infers one from the wire protocol. The provider prefix selects LiteLLM's request transformation and does not rename the actual channel. LiteLLM selects the default conversion behavior from the provider and model. DeepSeek and custom OpenAI-compatible channels can use native Responses after a direct probe passes and an administrator or upstream owner enables it.
Generic OpenAI-compatible channels use LiteLLM's custom_openai provider so Responses, Anthropic Messages, and Gemini requests can use LiteLLM's standard conversion path instead of being treated as native OpenAI endpoints. Xiaomi MiMo deliberately uses the openai provider because its current service exposes compatible Chat, Responses, and Anthropic APIs. MiniMax uses custom_openai with LiteLLM v1.102.0-rc.1 because that release's native minimax provider still cannot construct both the Chat and Anthropic endpoints from one deployment Base URL.
LiteLLM v1.102.0-rc.1 adds a native Responses adapter for Fireworks AI. This channel uses its own native provider, so Responses requests now default to the upstream native endpoint without the explicit switch used for DeepSeek and custom channels. The adapter does not support native WebSocket or File Search; validate HTTP Responses with the actual model during the upgrade.
Native Responses reports show separate HTTP, SSE, tool, and usage checks. A successful report remains valid while the target identity and runtime version stay unchanged, with retesting available when needed. Changes to the model, base URL, credential version, or runtime version require a new probe and manual enablement. The management page shows the saved configuration and database synchronization status: no deployment, pending, synchronized, or failed. Synchronized configuration does not establish the runtime’s current live state. The page refreshes when opened, after an action, or on manual refresh.
Mainland China and general upstreams
| Upstream | Protocol | Default base URL |
|---|---|---|
| DeepSeek | OpenAI-compatible | https://api.deepseek.com/v1 |
| Qwen | OpenAI-compatible | https://dashscope.aliyuncs.com/compatible-mode/v1 |
| GLM | OpenAI-compatible | https://open.bigmodel.cn/api/paas/v4 |
| Kimi | OpenAI-compatible | https://api.moonshot.cn/v1 |
| Doubao | OpenAI-compatible | https://ark.cn-beijing.volces.com/api/v3 |
| TokenHub | OpenAI-compatible | https://tokenhub.tencentmaas.com/v1 |
| MiniMax | OpenAI-compatible | https://api.minimaxi.com/v1 |
| StepFun | OpenAI-compatible | https://api.stepfun.com/v1 |
| MiMo | OpenAI-compatible | https://api.xiaomimimo.com/v1 |
| SiliconFlow | OpenAI-compatible | https://api.siliconflow.cn/v1 |
| Qianfan | OpenAI-compatible | https://qianfan.baidubce.com/v2 |
| Baichuan | OpenAI-compatible | https://api.baichuan-ai.com/v1 |
| Yi | OpenAI-compatible | https://api.lingyiwanwu.com/v1 |
| ModelScope | OpenAI-compatible | https://api-inference.modelscope.cn/v1 |
International upstreams
| Upstream | Protocol | Default base URL |
|---|---|---|
| OpenAI | OpenAI-compatible | https://api.openai.com/v1 |
| OpenRouter | OpenAI-compatible | https://openrouter.ai/api/v1 |
| Anthropic | Anthropic | https://api.anthropic.com |
| Gemini | Gemini | https://generativelanguage.googleapis.com |
| Cohere | Cohere | https://api.cohere.com/v2/chat |
| Mistral AI | OpenAI-compatible | https://api.mistral.ai/v1 |
| Groq | OpenAI-compatible | https://api.groq.com/openai/v1 |
| Together AI | OpenAI-compatible | https://api.together.xyz/v1 |
| Fireworks AI | OpenAI-compatible | https://api.fireworks.ai/inference/v1 |
| Cerebras | OpenAI-compatible | https://api.cerebras.ai/v1 |
| NVIDIA NIM | OpenAI-compatible | https://integrate.api.nvidia.com/v1 |
| Perplexity | OpenAI-compatible | https://api.perplexity.ai |
| Novita AI | OpenAI-compatible | https://api.novita.ai/openai/v1 |
| DeepInfra | OpenAI-compatible | https://api.deepinfra.com/v1/openai |
| SambaNova | OpenAI-compatible | https://api.sambanova.ai/v1 |
| xAI | OpenAI-compatible | https://api.x.ai/v1 |
Local and custom
| Upstream | Protocol | Default base URL |
|---|---|---|
| Ollama | Ollama | https://ollama.example.com |
| LM Studio | OpenAI-compatible | https://lm-studio.example.com/v1 |
| Custom OpenAI-compatible | OpenAI-compatible | User supplied |
In a container deployment, localhost points to the Lumoswitch container. Use a host address reachable from the container for Ollama or LM Studio. See Local models.
Verify actual capability
- Test the base URL and credential.
- Discover models visible to the account.
- Send a minimal chat request to the intended model.
- Test streaming, tools, vision, reasoning, or embeddings from the target client.
A preset does not grant access to every model or capability. Provider regions, projects, model IDs, and proprietary parameters can differ. OpenAI-compatible does not mean every OpenAI-specific parameter is supported.