Baseten

Baseten Model APIs provide hosted, OpenAI-compatible access to frontier models. The official external plugin uses authenticated discovery, so OpenClaw follows the complete model set enabled for your Baseten account. Its offline fallback contains the current curated models listed below.

Install plugin

openclaw plugins install @openclaw/baseten-provider

Installation applies to a running Gateway automatically; otherwise it takes effect on the next startup. See Apply changes and inspect.

Getting started

openclaw onboard --auth-choice baseten-api-key

openclaw onboard --non-interactive --accept-risk --skip-health \  --auth-choice baseten-api-key \  --baseten-api-key "$BASETEN_API_KEY"

export BASETEN_API_KEY=...

openclaw models list --provider baseten

Default model

DeepSeek V4.1 Flash is the default for new Baseten setups. It supports text and image input, tool calling, a 1,048,576-token context window, and up to 262,144 output tokens:

{  agents: {    defaults: {      model: { primary: "baseten/deepseek-ai/DeepSeek-V4.1-Flash" },    },  },}

Use /model baseten/deepseek-ai/DeepSeek-V4.1-Flash -s to switch the current session. Its thinking choices are off, low, high, and max, with high as the native default. OpenClaw sends off as reasoning_effort: "none". See Baseten's reasoning controls.

Saved explicit selections such as baseten/thinkingmachines/inkling are not redirected to the new default. If a previously selected model is no longer available, list the current catalog and choose a replacement explicitly. Reconnecting the API key preserves the saved primary model.

Bundled fallback catalog

The authenticated live catalog is authoritative. This curated fallback covers current Model APIs; live discovery can include additional models. Removing a retired row from this fallback does not delete an explicitly authored row from your config:

All bundled models support tool calling and reasoning. OpenClaw maps its thinking levels to models with native reasoning_effort. Its opt-in GLM and Nemotron routes default to thinking off. Nemotron exposes a binary off/on control; the bundled GLM 5.2 definitions expose off, high, and max. These routes use Baseten's chat_template_args.enable_thinking control. GLM's scalar effort and image input depend on discovery advertising those capabilities; sparse live GLM rows currently omit them.

The same thinking controls apply to agent turns and standalone model completions. Current DeepSeek V4 Pro 0813 and explicitly configured older Pro references preserve replayed reasoning metadata while thinking is enabled and remove it for explicit off requests. The current Pro endpoint additionally receives the required thinking envelope when scalar effort is sent.

Prompt caching

For the native https://inference.baseten.co/v1 endpoint, OpenClaw sends a stable x-session-affinity header for related requests to improve prefix cache reuse. It uses an explicit request cache key when present, otherwise the session ID. An explicitly configured affinity header takes precedence. Setting cacheRetention: "none" suppresses the generated header; it does not disable Baseten's automatic cache. Custom endpoints retain their configured behavior. See Baseten session affinity.

Manual config

Most setups only need the API key. To pin the provider explicitly:

{  env: { vars: { BASETEN_API_KEY: "..." } },  agents: {    defaults: {      model: { primary: "baseten/deepseek-ai/DeepSeek-V4.1-Flash" },    },  },  models: {    mode: "merge",    providers: {      baseten: {        baseUrl: "https://inference.baseten.co/v1",        apiKey: "${BASETEN_API_KEY}",        api: "openai-completions",        models: [          {            id: "deepseek-ai/DeepSeek-V4.1-Flash",            name: "DeepSeek V4.1 Flash",            reasoning: true,            input: ["text", "image"],            contextWindow: 1048576,            maxTokens: 262144,          },        ],      },    },  },}

Related