Nous ResearchObsidian
Obsidian 3B v0.5
Nous Research and Virtual Interactive built Obsidian from Capybara 3B v1.9, which uses StableLM 3B.
Preview
Training and model behavior
The model follows a LLaVA 1.5 training procedure. It uses a modified ChatML format with a separator. The repository links separate training and inference code.Obsidian 3B v0.5 model card
Working with this model
Practical guidance. Use images and questions that require visible evidence. Check the expected image preprocessing and chat separators before comparing responses.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Preview |
| Launch date | |
| Access | Downloadable |
| Model identifiers | NousResearch/Obsidian-3B-V0.5 |
| Context window | 4,096 tokens Configured position limit; practical serving limits depend on the runtime. Obsidian 3B v0.5 configuration |
| Base model | Capybara 3B v1.9 / StableLM-3B-4e1tObsidian 3B v0.5 model card |
| Repository license label | cc-by-sa-4.0Obsidian 3B v0.5 model card |