π§ Steerling-8B Demo
Steerling-8B is an 8 billion parameter causal diffusion language model with interpretable concept steering, built by Guide Labs.
Unlike standard autoregressive LLMs, Steerling generates text by iteratively unmasking tokens in order of confidence β the model fills in positions where it is most certain first. Watch the diffusion process live below!
β¨ Key Features
| Feature | Description |
|---|---|
| π² Diffusion decoding | Confidence-based unmasking instead of left-to-right |
| π Interpretability | Hidden states β known + unknown concept decomposition |
| ποΈ Concept steering | Amplify or suppress concepts to guide generation |
| π Block-causal attention | Bidirectional within 64-token blocks, causal across |
βΉοΈ This Space runs on ZeroGPU (NVIDIA H200). Generation may be queued briefly while a GPU is allocated.
Example prompts
How It Works
hidden β known_features + unknown_features + Ξ΅ = composed β logits
- known_features β weighted sum of top-k learned concept embeddings (interpretable)
- unknown_features β residual captured by a factorized unknown head
- Ξ΅ β small correction for reconstruction fidelity
The live visualization above shows the diffusion process in action:
- Blue text = your prompt
- Highlighted = just unmasked this step
- β = still masked (waiting to be filled)
Unlike autoregressive models that generate left-to-right, Steerling fills in the most confident positions first, regardless of order.
Links
- π Model Card
- π» GitHub
- π’ Guide Labs
- π Architecture Blog Post