Multi-Model Support

Hebb Mind supports multiple LLM providers through LiteLLM. This enables memory consolidation with any major language model.

Supported Providers

Provider	Model Example	Configuration
OpenAI	openai/gpt-4o-mini	llm_api_key
OpenAI	openai/gpt-4o	llm_api_key
Anthropic	anthropic/claude-3-haiku-20240307	llm_api_key
Anthropic	anthropic/claude-3-5-sonnet-20241022	llm_api_key
Qwen (Alibaba)	openai/qwen-plus	llm_api_key + llm_base_url
GLM (Zhipu)	openai/glm-4	llm_api_key + llm_base_url
Kimi (Moonshot)	openai/moonshot-v1-8k	llm_api_key + llm_base_url

Configuration

OpenAI

hebb config set llm_model openai/gpt-4o-mini
hebb config set llm_api_key sk-your-openai-key

Anthropic

hebb config set llm_model anthropic/claude-3-haiku-20240307
hebb config set llm_api_key sk-ant-your-anthropic-key

Qwen (Alibaba Cloud)

hebb config set llm_model openai/qwen-plus
hebb config set llm_api_key sk-your-qwen-key
hebb config set llm_base_url https://dashscope.aliyuncs.com/compatible-mode/v1

GLM (Zhipu AI)

hebb config set llm_model openai/glm-4
hebb config set llm_api_key your-zhipu-key
hebb config set llm_base_url https://open.bigmodel.cn/api/paas/v4

Kimi (Moonshot AI)

hebb config set llm_model openai/moonshot-v1-8k
hebb config set llm_api_key sk-your-moonshot-key
hebb config set llm_base_url https://api.moonshot.cn/v1

How It Works

For Chinese model providers (Qwen, GLM, Kimi), the openai/ prefix tells LiteLLM to use the OpenAI-compatible API format. The llm_base_url points to the provider's endpoint. This works because these providers implement the OpenAI chat completion API specification.

Embedding Model

The embedding model runs locally via sentence-transformers. No external API calls are needed for generating embeddings. By default, hebb setup selects a small model by content language:

• English: all-MiniLM-L6-v2 (~90 MB)
• Chinese or multilingual: intfloat/multilingual-e5-small (~470 MB)

For the high-quality tier (1–2 GB), opt in with hebb setup --profile best:

• English: BAAI/bge-large-en-v1.5
• Chinese or multilingual: BAAI/bge-m3

Download region is independent from language. Use hebb setup --language en --region cn for English content on a China network, or hebb setup --language zh --region global for Chinese content on a global network.

To swap to a different model after first install:

hebb config set embedding_model "paraphrase-multilingual-MiniLM-L12-v2"
hebb config set embedding_dim 384
hebb service restart
hebb memory reembed         # required if the dimension changed

TIP

Changing the embedding dimension invalidates all stored vectors — the vector table is auto-reset on next startup. Run hebb memory reembed afterwards to repopulate. See Switch the Embedding Model for the full walkthrough (CLI, Web Console, and re-embed details).

This means:

• Embedding is free -- no API costs
• Low latency -- no network round-trip
• Privacy -- your text never leaves the machine for embedding
• Offline capable -- works without internet after initial download

The embedding model is separate from the LLM model. You can use any LLM provider while keeping the local embedding model.

Testing Your Configuration

After configuring a model, test the connection:

curl -X POST http://localhost:8321/api/v1/admin/config/test-llm \
  -H 'Content-Type: application/json' \
  -d '{"model": "openai/gpt-4o-mini", "api_key": "sk-..."}'

This sends a simple test request to verify that the model, API key, and endpoint are working correctly. Add "base_url": "https://..." for OpenAI-compatible providers (Qwen, GLM, Kimi).

Choosing a Model

For memory consolidation, smaller and faster models work well since the task involves classification and summarization rather than creative generation. Recommended starting points:

• Budget-conscious: openai/gpt-4o-mini or openai/qwen-plus
• Higher quality: openai/gpt-4o or anthropic/claude-3-5-sonnet-20241022
• Chinese-language memories: openai/qwen-plus or openai/glm-4