The Model
Direct access to TasteBrain weights and the low-level multi-vector interface for advanced developers and researchers.
Overview
Level 0 provides the foundational model layer of TasteBrain. This level is designed for:
- Advanced developers and researchers who download models from Hugging Face
- Teams running local inference or operating directly on raw vectors
- Scientific exploration of high-dimensional "taste vectors"
Core Utility
Craft custom late-binding queries and experiment with high-dimensional "taste vectors" for scientific or proprietary style analysis.
Model Architecture
TasteBrain is built on Qwen3, a multi-modal foundation model, trained with contrastive learning on our proprietary dataset of aesthetic attributes.
Key Features
- Multi-Modal Embeddings - Unified embedding space for images, text, and products
- Aesthetic Manifold - High-dimensional space encoding "DNA of taste"
- Late-Binding Queries - Flexible multi-vector matching at inference time
- Cross-Domain Search - Map audio to products, food preferences to products, etc.
Deployment
Model cards include structured metadata for:
- Task identification - e.g., aesthetic-understanding, style-transfer
- Base-model tracking - Version control and lineage
- Performance benchmarks - Accuracy metrics on curated test sets
Current Usage
Level 0 is currently used exclusively by internal teams at Bestomer for:
- Research and development of new aesthetic understanding capabilities
- Direct access to vectors for operating on manifolds
- Custom query optimization and experimentation
- Training pipeline development
⚠️ Limited Public Access
Level 0 access is restricted to research partners and advanced use cases. For most developers, Level 1: SVK Primitives provides the right abstraction layer.
Next Steps
- Model Architecture - Deep dive into TasteBrain's design
- Taste Vectors - Understanding the embedding space
- Level 1: SVK Primitives - Production-ready APIs