Base models, served for research.
Direct, token-level access to large language base models, over an OpenAI-compatible API. Free for researchers.
We aim to support research in AI safety and interpretability, understanding LLM psychology and personas, cyborgism-style human-AI collaboration, and more.
What you get
Base models
You get the pretrained model, with no chat or instruction tuning. It reads your prompt exactly as you wrote it, with no template and no system prompt.
Token-level access
See the log probability of every token and of its top alternatives. For your prompt, you can get the full vocabulary at each position.
Activations & steering
Read the activations at the layers you choose, or add steering vectors to change what the model writes.
Playground and API
Try an idea in the browser Playground. Then run the full experiment through the OpenAI-compatible API.
Models
- meta-llama/
Llama-3.1-8B - meta-llama/
Llama-3.1-405B - arcee-ai/
Trinity-Large-TrueBase