Configure the LLM provider

Read this if you want to change which model OMem uses, switch from a subscription to an API key, or point OMem at a local model. OMem doesn’t ship a model — you bring your own — and the curator (text → wiki page) and VLM (image → description) are configured independently.

The four providers

anthropic-oauth (the default) — your Claude Pro / Max subscription. No API key. Set it up through the wizard:

omem setup --llm

anthropic-api — billed per token. Store the key in an env var or the Keychain (never in the config file):

omem config set llm.curator.provider anthropic-api
omem config set llm.curator.model claude-sonnet-4-6
omem config set llm.curator.api_key_env ANTHROPIC_API_KEY

openai-compat — any endpoint that speaks the OpenAI API: OpenAI, Chinese models (Alibaba Bailian / ByteDance Volcengine Ark / Kimi / Zhipu GLM / MiniMax / DeepSeek, etc.), a hosted gateway, or a local server. The escape hatch:

omem config set llm.curator.provider openai-compat
omem config set llm.curator.base_url http://localhost:11434/v1
omem config set llm.curator.model your-model-name
omem config set llm.curator.api_key_env YOUR_KEY_ENV

For exact base_url and model names for Chinese model platforms, see Using Chinese model platforms below.

openai-chatgpt-oauth — your ChatGPT Plus / Pro subscription. Like the Claude option, the wizard handles sign-in:

omem setup --llm

Curator and VLM are separate

You can use a strong text model for curation and a cheaper (or different-vendor) vision model for images — they’re independent under llm.curator.* and llm.vlm.*:

omem config set llm.vlm.provider openai-compat
omem config set llm.vlm.model qwen-vl-max

Using Chinese model platforms

The major Chinese model platforms all expose OpenAI-compatible endpoints — point openai-compat at the right base_url and use the platform’s model ID:

Platform	`base_url`	Curator model	Vision model
Alibaba Bailian (Qwen)	`https://dashscope.aliyuncs.com/compatible-mode/v1`	`qwen-max` / `qwen-plus`	`qwen-vl-max`
ByteDance Volcengine Ark (Doubao)	`https://ark.cn-beijing.volces.com/api/v3`	`doubao-1-5-pro-32k`	`doubao-1-5-vision-pro`
Kimi (Moonshot)	`https://api.moonshot.cn/v1`	`kimi-k2-0905-preview`	`moonshot-v1-32k-vision-preview`
Zhipu GLM	`https://open.bigmodel.cn/api/paas/v4`	`glm-4-plus`	`glm-4v-plus`
MiniMax	`https://api.minimax.chat/v1`	`MiniMax-Text-01`	(check official vision model)
DeepSeek	`https://api.deepseek.com/v1`	`deepseek-chat`	(no first-party vision yet; use another provider for VLM)

Example — Bailian for both curator (qwen-max) and VLM (qwen-vl-max):

# Curator
omem config set llm.curator.provider openai-compat
omem config set llm.curator.base_url https://dashscope.aliyuncs.com/compatible-mode/v1
omem config set llm.curator.model qwen-max
omem config set llm.curator.api_key_env DASHSCOPE_API_KEY

# Vision model (can be a different provider; here we keep it on Bailian)
omem config set llm.vlm.provider openai-compat
omem config set llm.vlm.base_url https://dashscope.aliyuncs.com/compatible-mode/v1
omem config set llm.vlm.model qwen-vl-max
omem config set llm.vlm.api_key_env DASHSCOPE_API_KEY

# Set the key in your shell (or via Keychain — see above)
export DASHSCOPE_API_KEY=sk-xxx

Always check the platform’s official docs for the latest base_url and model IDs — they ship new models and deprecate old ones often.

Running a local model

Set openai-compat and point base_url at your local server (Ollama, vLLM, LM Studio, …). It works today, with one honest caveat: quality depends on the model. OMem’s “good by default” promise is calibrated to capable cloud models; a small local model will curate, but with lower fidelity. Fully-offline, work-grade quality is a v1.5 goal.

Verify

omem doctor

The llm endpoint check confirms your provider is reachable and configured. See the full field list in the config schema.

What’s next

Query and search — once pages are curated, find them.
Configuration schema → llm — every provider field.