> ## Documentation Index > Fetch the complete documentation index at: https://docs.loongclaw.ai/llms.txt > Use this file to discover all available pages before exploring further. # Providers And Models > Choose a provider, understand model selection, and keep provider truth explicit. # Providers And Models Use this page when the base CLI path already makes sense and you need to choose the right provider lane before editing config in detail. Loong treats provider and model choice as operator-visible configuration, not hidden runtime magic. This page is the chooser. The exact built-in provider contract lives in [Provider Guides](/use-loong/provider-guides/index), while [Provider Recipes](/use-loong/provider-recipes) keeps a smaller set of representative rollout patterns. Volcengine and BytePlus stay near the top because they are a common first-run path, but they are not the only provider families surfaced here. If you want the shared public provider-profile shape before the recipes become provider-specific, start with [Configuration Patterns](/use-loong/configuration-patterns). If you already know the provider choice and delivery surface together, go straight to [Common Setups](/use-loong/common-setups) and the dedicated playbook that matches the rollout shape. ## Choose A Provider Lane | If you need... | Start here | Why | | --------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | | a common hosted first run with a short reviewed path | Volcengine or BytePlus | the main hosted fast path stays short without hiding the broader provider map | | a direct hosted provider your team already uses | OpenAI, Anthropic, Gemini, DeepSeek, MiniMax, Mistral, Cohere, xAI, Z.ai, Zhipu, Qwen, Kimi, Groq, Fireworks, Together, Perplexity, Qianfan, Cerebras, NVIDIA, SambaNova, Novita, SiliconFlow, StepFun, or Venice | built-in families stay explicit instead of being forced through `custom` | | a separate coding lane for coding quota or billing | Volcengine Coding, BytePlus Coding, Kimi Coding, Step Plan, or Bailian Coding | coding traffic stays separate from the general hosted lane | | one proxy or compatibility endpoint in front of upstream models | gateway or custom endpoint recipes | route ownership stays visible instead of being mistaken for a first-party provider | | a local or self-hosted model server | Ollama, LM Studio, llama.cpp, vLLM, or SGLang | the server boundary and model inventory stay explicit | | several reviewed provider profiles in one config | multi-profile provider recipes | cost, quota, latency, or region decisions stay operator-visible | ## Already Know The Whole Rollout? If the provider choice is already tied to the delivery surface, skip this page and jump straight to the matching playbook. | If you already know you want... | Go here | | --------------------------------------------------- | ----------------------------------------------------------------------------- | | Volcengine plus Feishu or Lark | [Volcengine Plus Feishu Or Lark](/use-loong/volcengine-feishu-lark-playbook) | | Volcengine plus WeCom | [WeCom Rollout](/use-loong/wecom-rollout-playbook) | | BytePlus Coding plus Telegram | [BytePlus Coding Plus Telegram](/use-loong/byteplus-coding-telegram-playbook) | | a local or self-hosted model plus outbound delivery | [Local Model Plus Outbound Delivery](/use-loong/local-outbound-playbook) | ## Start With The Supported Path ```bash theme={null} loong # If config is missing, first-run setup stays inside the main shell loong ask --message "Summarize this repository and suggest the best next step." loong runtime models list ``` The main `loong` shell owns the supported first-run path. When config is missing, the guided setup stays inside that same shell and hands you off to the same provider surface the runtime will keep using later. Use `loong onboard` when you want to reopen the deeper guided setup flow explicitly for provider or channel reconfiguration. Use `runtime models list` after credentials are healthy when you want the current catalog from the active provider. ## Reading Rule * Use this page to choose the provider lane. * Use [Configuration Patterns](/use-loong/configuration-patterns) when you want the shared `active_provider` plus `providers.` shape before the provider-specific walkthroughs. * Use [Provider Guides](/use-loong/provider-guides/index) once you know the lane and need the exact built-in contract. * Use [Provider Recipes](/use-loong/provider-recipes) when you want representative rollout patterns on top of that contract. * Use [Common Setups](/use-loong/common-setups) when the provider choice only makes sense together with the delivery surface. ## How Model Selection Works | Mode | What it means | When to use it | | ----------------------- | ---------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------- | | Explicit model | `provider.model = "..."` pins one model name | when you want deterministic behavior or already know the exact model | | Auto discovery | `provider.model = "auto"` asks the runtime to discover usable models from the provider catalog | when you want a clean first-run path and the provider exposes a model catalog | | Preferred fallback list | `preferred_models = ["...", "..."]` defines an explicit operator fallback order | when catalog discovery is unstable or you want a reviewed fallback without inventing a hidden default | * `preferred_models` is only an operator-configured fallback list. * It is not a secret provider-owned runtime default. * When the config already pins an explicit model, `preferred_models` does not override it. ## Reviewed Onboarding Defaults A small reviewed subset currently has an explicit onboarding fallback when the saved config still uses `model = "auto"` and model discovery is unavailable. | Provider | Reviewed onboarding default | | -------- | --------------------------- | | DeepSeek | `deepseek-chat` | | MiniMax | `MiniMax-M2.7` | For other providers, Loong prefers to keep `auto` or ask the operator for an explicit model instead of silently inventing one. ## Built-In Provider Map You do not need to memorize every built-in provider kind to get started. Start with the lane chooser above, then use this section as the broader public map. A practical public grouping, inferred from the built-in provider kinds, is: ### Hosted Families At A Glance | Family | Current built-in kinds | Use it when... | | ---------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- | | platform-first hosted | Volcengine, BytePlus | you want a short first-run hosted lane that stays explicit in the docs | | direct hosted general | OpenAI, Anthropic, Gemini, DeepSeek, MiniMax, Mistral, Cohere, xAI, Z.ai, Zhipu, Qwen, Kimi, Groq, Fireworks, Together, Perplexity, Qianfan, Cerebras, NVIDIA, SambaNova, Novita, SiliconFlow, StepFun, Venice | your team already has a named hosted provider and you want that family to remain visible | | coding-specialized hosted | Volcengine Coding, BytePlus Coding, Kimi Coding, Step Plan, Bailian Coding | coding work, quota, or pricing should stay on an explicit lane | | gateway and platform routing | OpenRouter, Cloudflare AI Gateway, Vercel AI Gateway, Bedrock, Custom | a gateway, proxy, or managed route sits between Loong and the upstream models | | local and self-hosted | Ollama, LM Studio, llama.cpp, vLLM, SGLang | the model server boundary should stay operator-owned | ### Platform-First Hosted Providers * Volcengine * Volcengine Coding * BytePlus * BytePlus Coding ### Direct Hosted Providers * OpenAI * Anthropic * Gemini * DeepSeek * MiniMax * Mistral * Cohere * xAI * Z.ai * Zhipu * Qwen * Kimi * Groq * Fireworks * Together * Perplexity * Qianfan * Cerebras * NVIDIA * SambaNova * Novita * SiliconFlow * StepFun * Venice ### Coding-Specialized Hosted Providers * Kimi Coding * Step Plan * Bailian Coding ### Gateway And Compatibility Providers * Cloudflare AI Gateway * Vercel AI Gateway * OpenRouter * Bedrock * Custom OpenAI-compatible endpoints ### Local And Self-Hosted Providers * Ollama * LM Studio * llama.cpp * vLLM * SGLang The exact active set still depends on the build, the configured profile, and whether credentials resolve on the current machine. ## Minimal Config Shape The first-run path should normally write this for you, but the public config shape looks like this: ```toml theme={null} active_provider = "volcengine" [providers.volcengine] kind = "volcengine" api_key = { env = "ARK_API_KEY" } model = "auto" ``` The important part is that the runtime uses the same provider surface after onboarding that it uses during normal operation. There is no separate marketing-only provider layer. Here `api_key = { env = "ARK_API_KEY" }` explicitly means "read from the `ARK_API_KEY` environment variable". Writing `api_key = "ARK_API_KEY"` would instead treat `ARK_API_KEY` as the literal key value, which is why the docs do not use that form for env-backed secrets. The same profile shape also works for OpenAI, Anthropic, Gemini, DeepSeek, MiniMax, BytePlus, and other built-in hosted families by swapping the provider `kind`, the default credential env, and the reviewed model choice. ## Continue Reading * Continue to [Provider Guides](/use-loong/provider-guides/index) for the full built-in provider matrix. * Continue to [Provider Recipes](/use-loong/provider-recipes) for step-by-step hosted, coding-lane, gateway, local, and multi-profile examples. * Go to [First Run](/get-started/first-run) if you still need the shortest path to a useful answer. * Go to [Common Setups](/use-loong/common-setups) if the provider choice is already tied to Feishu / Lark, WeCom, Telegram, local outbound delivery, or gateway ownership. * Go to [Channels](/use-loong/channels) if you want the current delivery-surface model. * Go to [Tools And Memory](/use-loong/tools-and-memory) if you want the broader governed-capability story.